Eric Zietlow's blog covering Tenstorrent hardware, Metalium programming, and AI topics, sharing practical experience with Blackhole and the broader TT ecosystem from our developer relations team.
\n\nLinks: dev.to/mando222
\nBy Eric Zietlow · affiliated · added 2026-08-12
\nblog, metalium, blackhole, ai, dev-to, planet, blogs
", "summary": "Eric Zietlow's blog covering Tenstorrent hardware, Metalium programming, and AI topics, sharing practical experience with Blackhole and the broader TT ecosystem from our developer relations team.", "date_published": "2026-08-12T23:59:59Z", "tags": [ "blogs", "affiliated", "entry" ] }, { "id": "https://github.com/tenstorrent/tt-system-firmware/releases/tag/v19.13.2", "url": "https://github.com/tenstorrent/tt-system-firmware/releases/tag/v19.13.2", "title": "tt-system-firmware v19.13.2", "content_html": "Blackhole boards can now harvest individual GDDR channels per instance through a new firmware table field that's configurable at runtime, giving you finer control over memory defect handling without reflashing. Beyond that, v19.13.1–19.13.2 bring telemetry improvements across platforms (feature capability reporting on Blackhole, Metal runtime telemetry carve-outs on Wormhole), a new KMD logging backend for streaming firmware logs over PCIe, and a stability fix for intermittent Tensix reset hangs by switching to NOC coordinates; check the migration guide when upgrading from v19.12.0.
\n\nLinks: Repository · v19.13.2
\nofficial · added 2026-08-11
", "summary": "Blackhole boards can now harvest individual GDDR channels per instance through a new firmware table field that's configurable at runtime, giving you finer control over memory defect handling without reflashing. Beyond that, v19.13.1–19.13.2 bring telemetry improvements across platforms (feature capability reporting on Blackhole, Metal runtime telemetry carve-outs on Wormhole), a new [KMD logging backend](https://github.com/tenstorrent/tt-system-firmware/tree/main/doc/services/kmd_logging/index.rst) for streaming firmware logs over PCIe, and a stability fix for intermittent Tensix reset hangs by switching to NOC coordinates; check the [migration guide](https://github.com/tenstorrent/tt-system-firmware/tree/main/doc/release/migration-guide-19.13.md) when upgrading from v19.12.0.", "date_published": "2026-08-11T21:11:11Z", "tags": [ "official", "release" ] }, { "id": "https://github.com/tenstorrent/tt-toplike/releases/tag/v0.8.0", "url": "https://github.com/tenstorrent/tt-toplike/releases/tag/v0.8.0", "title": "tt-toplike v0.8.0", "content_html": "tt-toplike now exposes the full span of driver telemetry — including tt-kmd class attributes, live PCIe bandwidth, and tt-smi 6.x metrics — while fixing a bug where per-device data could shuffle between reads, making monitoring and profiling workflows far more reliable for developers instrumenting multi-device clusters.
\n\nLinks: Repository · v0.8.0
\nofficial · added 2026-08-11
", "summary": "tt-toplike now exposes the full span of driver telemetry — including tt-kmd class attributes, live PCIe bandwidth, and tt-smi 6.x metrics — while fixing a bug where per-device data could shuffle between reads, making monitoring and profiling workflows far more reliable for developers instrumenting multi-device clusters.", "date_published": "2026-08-11T20:30:47Z", "tags": [ "official", "release" ] }, { "id": "https://github.com/tenstorrent/tt-smi/releases/tag/v6.2.0", "url": "https://github.com/tenstorrent/tt-smi/releases/tag/v6.2.0", "title": "tt-smi v6.2.0", "content_html": "Total board power is now visible in tt-smi's telemetry table and snapshot output, giving you a unified view of your hardware's power consumption without hunting through per-chip metrics. This matters for capacity planning and understanding thermal headroom across your Tenstorrent clusters at a glance.
Links: Repository · v6.2.0
\nofficial · added 2026-08-11
", "summary": "Total board power is now visible in `tt-smi`'s telemetry table and snapshot output, giving you a unified view of your hardware's power consumption without hunting through per-chip metrics. This matters for capacity planning and understanding thermal headroom across your Tenstorrent clusters at a glance.", "date_published": "2026-08-11T15:40:06Z", "tags": [ "official", "release" ] }, { "id": "https://github.com/tenstorrent/tt-metal/releases/tag/v0.76.0", "url": "https://github.com/tenstorrent/tt-metal/releases/tag/v0.76.0", "title": "tt-metal v0.76.0", "content_html": "This release brings significant stability and feature improvements across low-level kernel execution, the core Metal runtime, and high-level neural network operations. Most notably, Quasar support now handles dual command queues, while a comprehensive sweep of data-race fixes in multicast and dataflow operations closes long-standing correctness gaps in distributed workloads—particularly in collective communication ops and dynamic tensor operations. On the kernel side, SFPU consolidation (#50399), improved numerical precision for transcendental functions, and broader Quasar performance test coverage reflect maturing silicon support. For developers building models, the Metal 2.0 migration continues accelerating (15+ factory ports completed) and per-core memory allocation now has a unified API (#51500), reducing complexity in sharded tensor workflows. A large batch of infrastructure improvements—from in-place wheel builds speeding up CI to agentic workflows automating triage—round out a release focused on correctness, performance, and developer velocity.
\n\nLinks: Repository · v0.76.0
\nofficial · added 2026-08-11
", "summary": "This release brings significant stability and feature improvements across low-level kernel execution, the core Metal runtime, and high-level neural network operations. Most notably, [Quasar support](https://github.com/tenstorrent/tt-metal/pull/50280) now handles dual command queues, while a comprehensive sweep of data-race fixes in multicast and dataflow operations closes long-standing correctness gaps in distributed workloads—particularly in collective communication ops and dynamic tensor operations. On the kernel side, SFPU consolidation ([#50399](https://github.com/tenstorrent/tt-metal/pull/50399)), improved numerical precision for transcendental functions, and broader Quasar performance test coverage reflect maturing silicon support. For developers building models, the Metal 2.0 migration continues accelerating (15+ factory ports completed) and per-core memory allocation now has a unified API ([#51500](https://github.com/tenstorrent/tt-metal/pull/51500)), reducing complexity in sharded tensor workflows. A large batch of infrastructure improvements—from in-place wheel builds speeding up CI to agentic workflows automating triage—round out a release focused on correctness, performance, and developer velocity.", "date_published": "2026-08-11T04:10:19Z", "tags": [ "official", "release" ] }, { "id": "https://github.com/prag79/tt-sim-lab", "url": "https://github.com/prag79/tt-sim-lab", "title": "tt-sim Lab", "content_html": "A university teaching lab for TT-Metalium kernel programming on a virtual Tenstorrent chip — one-click GitHub Codespace, no silicon and nothing installed locally. The primary track (labs 00-06) points tt-metal straight at libttsim via TT_METAL_SIMULATOR and walks from elementwise add through NoC multicast to multi-core and multicast matmul, backed by a source-level matmul guide. An optional advanced track (labs 10-16) boots an Ubuntu guest under ttsim-qemu, loads tt-kmd, surfaces /dev/tenstorrent/0, and runs tt-metal through the full PCIe path.
\n\nLinks: Repo · Student handout · Matmul source walkthrough
\nBy @Pragnajit Datta Roy · affiliated · added 2026-08-10
\nttsim, simulator, qemu, codespaces, metalium, matmul, labs, curriculum, academia, education, guides, kernels, getting-started
", "summary": "A university teaching lab for TT-Metalium kernel programming on a virtual Tenstorrent chip — one-click GitHub Codespace, no silicon and nothing installed locally. The primary track (labs 00-06) points tt-metal straight at libttsim via TT_METAL_SIMULATOR and walks from elementwise add through NoC multicast to multi-core and multicast matmul, backed by a source-level matmul guide. An optional advanced track (labs 10-16) boots an Ubuntu guest under ttsim-qemu, loads tt-kmd, surfaces /dev/tenstorrent/0, and runs tt-metal through the full PCIe path.", "date_published": "2026-08-10T23:59:58Z", "tags": [ "guides", "kernels", "getting-started", "affiliated", "entry" ] }, { "id": "https://github.com/prag79/tt-sim-lab/blob/main/HANDOUT.md", "url": "https://github.com/prag79/tt-sim-lab/blob/main/HANDOUT.md", "title": "tt-sim Lab — Student handout", "content_html": "A university teaching lab for TT-Metalium kernel programming on a virtual Tenstorrent chip — one-click GitHub Codespace, no silicon and nothing installed locally. The primary track (labs 00-06) points tt-metal straight at libttsim via TT_METAL_SIMULATOR and walks from elementwise add through NoC multicast to multi-core and multicast matmul, backed by a source-level matmul guide. An optional advanced track (labs 10-16) boots an Ubuntu guest under ttsim-qemu, loads tt-kmd, surfaces /dev/tenstorrent/0, and runs tt-metal through the full PCIe path.
\n\nLinks: Student handout
\nBy @Pragnajit Datta Roy · affiliated · added 2026-08-10
\nttsim, simulator, qemu, codespaces, metalium, matmul, labs, curriculum, academia, education, guides, kernels, getting-started
", "summary": "A university teaching lab for TT-Metalium kernel programming on a virtual Tenstorrent chip — one-click GitHub Codespace, no silicon and nothing installed locally. The primary track (labs 00-06) points tt-metal straight at libttsim via TT_METAL_SIMULATOR and walks from elementwise add through NoC multicast to multi-core and multicast matmul, backed by a source-level matmul guide. An optional advanced track (labs 10-16) boots an Ubuntu guest under ttsim-qemu, loads tt-kmd, surfaces /dev/tenstorrent/0, and runs tt-metal through the full PCIe path.", "date_published": "2026-08-10T23:59:58Z", "tags": [ "guides", "kernels", "getting-started", "affiliated", "lesson", "article" ] }, { "id": "https://github.com/prag79/tt-sim-lab/blob/main/labs/MATMUL_GUIDE.md", "url": "https://github.com/prag79/tt-sim-lab/blob/main/labs/MATMUL_GUIDE.md", "title": "tt-sim Lab — Matmul source walkthrough", "content_html": "A university teaching lab for TT-Metalium kernel programming on a virtual Tenstorrent chip — one-click GitHub Codespace, no silicon and nothing installed locally. The primary track (labs 00-06) points tt-metal straight at libttsim via TT_METAL_SIMULATOR and walks from elementwise add through NoC multicast to multi-core and multicast matmul, backed by a source-level matmul guide. An optional advanced track (labs 10-16) boots an Ubuntu guest under ttsim-qemu, loads tt-kmd, surfaces /dev/tenstorrent/0, and runs tt-metal through the full PCIe path.
\n\nLinks: Matmul source walkthrough
\nBy @Pragnajit Datta Roy · affiliated · added 2026-08-10
\nttsim, simulator, qemu, codespaces, metalium, matmul, labs, curriculum, academia, education, guides, kernels, getting-started
", "summary": "A university teaching lab for TT-Metalium kernel programming on a virtual Tenstorrent chip — one-click GitHub Codespace, no silicon and nothing installed locally. The primary track (labs 00-06) points tt-metal straight at libttsim via TT_METAL_SIMULATOR and walks from elementwise add through NoC multicast to multi-core and multicast matmul, backed by a source-level matmul guide. An optional advanced track (labs 10-16) boots an Ubuntu guest under ttsim-qemu, loads tt-kmd, surfaces /dev/tenstorrent/0, and runs tt-metal through the full PCIe path.", "date_published": "2026-08-10T23:59:58Z", "tags": [ "guides", "kernels", "getting-started", "affiliated", "lesson", "article" ] }, { "id": "https://github.com/aldegad/tenstorrent", "url": "https://github.com/aldegad/tenstorrent", "title": "Tenstorrent Console Skill", "content_html": "An agent skill (SKILL.md) that teaches Claude Code, Codex, and the Agent SDK how to drive the console.tenstorrent.com inference API: OpenAI-compatible chat with DeepSeek-R1 and Qwen3, async image jobs, and Wan 2.2 text-to-video. Ships runnable curl examples and a mock-curl test harness; documentation is bilingual Korean/English.
\n\nLinks: Repo
\nBy @aldegad · community · added 2026-08-10
\nagent-skill, claude-code, codex, agent-sdk, tenstorrent-console, inference-api, text-to-video, wan2-2, korean, agents, dev-tools
", "summary": "An agent skill (SKILL.md) that teaches Claude Code, Codex, and the Agent SDK how to drive the console.tenstorrent.com inference API: OpenAI-compatible chat with DeepSeek-R1 and Qwen3, async image jobs, and Wan 2.2 text-to-video. Ships runnable curl examples and a mock-curl test harness; documentation is bilingual Korean/English.", "date_published": "2026-08-10T23:59:57Z", "tags": [ "agents", "dev-tools", "community", "entry" ] }, { "id": "https://github.com/aldegad/tenstorrent-cli", "url": "https://github.com/aldegad/tenstorrent-cli", "title": "tenstorrent-cli", "content_html": "A TypeScript/Bun terminal client for console.tenstorrent.com. Opens a chat REPL across DeepSeek-R1, Qwen3-32B, Qwen3-VL, and Gemma, with slash commands that submit image, Wan 2.2 video, TTS, and STT jobs, poll them, and save the results under ./output. Reads its API key only from the TENSTORRENT_KEY environment variable — never from disk.
\n\nLinks: Repo
\nBy @aldegad · community · added 2026-08-10
\ncli, repl, typescript, bun, tenstorrent-console, inference-api, text-to-video, dev-tools
", "summary": "A TypeScript/Bun terminal client for console.tenstorrent.com. Opens a chat REPL across DeepSeek-R1, Qwen3-32B, Qwen3-VL, and Gemma, with slash commands that submit image, Wan 2.2 video, TTS, and STT jobs, poll them, and save the results under ./output. Reads its API key only from the TENSTORRENT_KEY environment variable — never from disk.", "date_published": "2026-08-10T23:59:56Z", "tags": [ "dev-tools", "community", "entry" ] }, { "id": "https://github.com/antonibertel/tt-monitor", "url": "https://github.com/antonibertel/tt-monitor", "title": "tt-monitor", "content_html": "A translucent, undecorated desktop widget showing live per-chip telemetry for Tenstorrent accelerators: temperature and power sparklines against the card's thermal and TDP limits, AI clock, voltage, current, DRAM channel training and ECC error counts, PCIe link generation/width, and board identity. Reads hardware directly through luwen over /dev/tenstorrent — no Python, no tt-smi subprocess, and no root.
\n\nLinks: Repo
\nBy @antonibertel · community · added 2026-08-10
\ntelemetry, monitoring, widget, luwen, gtk4, rust, ecc, desktop, dev-tools, hw-system
", "summary": "A translucent, undecorated desktop widget showing live per-chip telemetry for Tenstorrent accelerators: temperature and power sparklines against the card's thermal and TDP limits, AI clock, voltage, current, DRAM channel training and ECC error counts, PCIe link generation/width, and board identity. Reads hardware directly through luwen over /dev/tenstorrent — no Python, no tt-smi subprocess, and no root.", "date_published": "2026-08-10T23:59:55Z", "tags": [ "dev-tools", "hw-system", "community", "entry" ] }, { "id": "https://github.com/marty1885/ttWKV7", "url": "https://github.com/marty1885/ttWKV7", "title": "ttWKV7", "content_html": "A standalone tt-metal demo and test bench for the RWKV-7 (WKV7) state recurrence on Wormhole, built with a GGML backend in mind. Two compute kernels cover the domain: a chunked-parallel DPLR matmul path for any sequence length with on-chip chunk carry, and a sequential per-token decode path for L <= 32 that is faster for large-batch token generation. The host runner validates both against a CPU oracle by PCC/NMSE and benchmarks them over a sequence/batch grid.
\n\nLinks: Repo
\nBy @Martin Chang · community · added 2026-08-10
\nrwkv, wkv7, operator, tt-metal, ggml, recurrent, cpp, kernels
", "summary": "A standalone tt-metal demo and test bench for the RWKV-7 (WKV7) state recurrence on Wormhole, built with a GGML backend in mind. Two compute kernels cover the domain: a chunked-parallel DPLR matmul path for any sequence length with on-chip chunk carry, and a sequential per-token decode path for L <= 32 that is faster for large-batch token generation. The host runner validates both against a CPU oracle by PCC/NMSE and benchmarks them over a sequence/batch grid.", "date_published": "2026-08-10T23:59:54Z", "tags": [ "kernels", "community", "entry" ] }, { "id": "https://github.com/Gogopex/tt-tinygrad", "url": "https://github.com/Gogopex/tt-tinygrad", "title": "tt-tinygrad", "content_html": "A Tenstorrent backend for tinygrad that targets TT-Lang rather than raw tt-metal: a Renderer classifies each UOp kernel graph as matmul, reduce, or elementwise and emits ttl.math.* Python source, and a Compiled device parses the rendered kernel's contract, materializes ttnn tensors from host buffers, and calls it in-process. Proof of concept — 110 pass / 13 xfail across 125 cases on a QuietBox, covering fused matmul, reductions, softmax, layernorm, and attention chains, on top of a three-line patch to upstream tinygrad.
\n\nLinks: Repo
\nBy @Gogopex · community · added 2026-08-10
\ntinygrad, tt-lang, ttnn, backend, codegen, renderer, python, compilers
", "summary": "A Tenstorrent backend for tinygrad that targets TT-Lang rather than raw tt-metal: a Renderer classifies each UOp kernel graph as matmul, reduce, or elementwise and emits ttl.math.* Python source, and a Compiled device parses the rendered kernel's contract, materializes ttnn tensors from host buffers, and calls it in-process. Proof of concept — 110 pass / 13 xfail across 125 cases on a QuietBox, covering fused matmul, reductions, softmax, layernorm, and attention chains, on top of a three-line patch to upstream tinygrad.", "date_published": "2026-08-10T23:59:53Z", "tags": [ "compilers", "community", "entry" ] }, { "id": "https://github.com/OrangeTangy/tensix-field-guide", "url": "https://github.com/OrangeTangy/tensix-field-guide", "title": "Tensix Field Guide", "content_html": "A ten-chapter, plain-English tour of Tenstorrent's Tensix architecture written for someone who knows what a CPU and a GPU are and nothing else: the chip-level grid and NoC, the five RISC-V baby cores, the matrix engine and its LoFi/HiFi fidelity trade-off, the SFPU, L1 and circular buffers, and why everything is 32x32 tile-shaped. The goal is to get a newcomer to the point of reading tt-metal kernel code in one sitting. Self-described draft; every claim traces back to tt-metal tech reports, tt-llk docs, or METALIUM_GUIDE.
\n\nLinks: Repo
\nBy @Tanay Anand · community · added 2026-08-10
\ntensix, architecture, sfpu, noc, circular-buffers, tile-format, risc-v, education, book, guides, riscv-arch
", "summary": "A ten-chapter, plain-English tour of Tenstorrent's Tensix architecture written for someone who knows what a CPU and a GPU are and nothing else: the chip-level grid and NoC, the five RISC-V baby cores, the matrix engine and its LoFi/HiFi fidelity trade-off, the SFPU, L1 and circular buffers, and why everything is 32x32 tile-shaped. The goal is to get a newcomer to the point of reading tt-metal kernel code in one sitting. Self-described draft; every claim traces back to tt-metal tech reports, tt-llk docs, or METALIUM_GUIDE.", "date_published": "2026-08-10T23:59:52Z", "tags": [ "guides", "riscv-arch", "community", "entry" ] }, { "id": "https://github.com/Zaneham/Booth/releases/tag/v0.5.2", "url": "https://github.com/Zaneham/Booth/releases/tag/v0.5.2", "title": "BarraCUDA v0.5.2", "content_html": "Booth 0.5.2 fixes a series of silent failures where the compiler was cheerfully reporting success while the backend ran on rejected code, wrote output files over the same path, or lost precision without warning—problems that sat there for a while because they only showed up once compilation finished. The structural story is the backend contract: targets now register themselves in a single list behind a be_desc_t, which shrunk main.c by a third and made adding a new target a matter of reading one header instead of 420 KB of AMD reference code. On the distribution side, Booth now installs properly with make install, ships static binaries for Linux, macOS, and Windows with no runtime dependencies, and exposes a CMake package so downstream projects can call booth_add_kernel() as part of their own build, with a worked example under examples/cmake/ and CI validation on every push.
Links: Repository · v0.5.2
\ncommunity · added 2026-08-08
", "summary": "Booth 0.5.2 fixes a series of silent failures where the compiler was cheerfully reporting success while the backend ran on rejected code, wrote output files over the same path, or lost precision without warning—problems that [sat there](https://github.com/Zaneham/Booth/blob/master/CHANGELOG.md) for a while because they only showed up once compilation finished. The structural story is the backend contract: targets now register themselves in a single list behind a `be_desc_t`, which shrunk `main.c` by a third and made adding a new target a matter of reading one header instead of 420 KB of AMD reference code. On the distribution side, Booth now installs properly with `make install`, ships static binaries for Linux, macOS, and Windows with no runtime dependencies, and exposes a CMake package so downstream projects can call `booth_add_kernel()` as part of their own build, with a worked example under `examples/cmake/` and CI validation on every push.", "date_published": "2026-08-08T03:05:39Z", "tags": [ "community", "release" ] }, { "id": "https://github.com/tenstorrent/ttsim/releases/tag/v1.10.0", "url": "https://github.com/tenstorrent/ttsim/releases/tag/v1.10.0", "title": "ttsim v1.10.0", "content_html": "QSR Tensix simulation is getting substantially more complete in this release, with SFPSTORE, ZEROACC, and a handful of other core LLK operations now properly modeled across destination rows and clear modes—crucial if you're running real kernel code through ttsim and want the simulator to stay in sync with actual hardware behavior. You'll also find fresh support for BH QuietBox2 configurations, fixes to Ethernet state handling on WH, and better MMIO introspection, all aimed at closing the gap between simulation and metal.
\n\nLinks: Repository · v1.10.0
\nofficial · added 2026-08-07
", "summary": "QSR Tensix simulation is getting substantially more complete in this release, with [SFPSTORE](https://github.com/tenstorrent/ttsim/blob/main/), [ZEROACC](https://github.com/tenstorrent/ttsim/blob/main/), and a handful of other core LLK operations now properly modeled across destination rows and clear modes—crucial if you're running real kernel code through ttsim and want the simulator to stay in sync with actual hardware behavior. You'll also find fresh support for BH QuietBox2 configurations, fixes to Ethernet state handling on WH, and better MMIO introspection, all aimed at closing the gap between simulation and metal.", "date_published": "2026-08-07T21:33:47Z", "tags": [ "official", "release" ] }, { "id": "https://github.com/moritztng/tt-bio/releases/tag/v0.6.2", "url": "https://github.com/moritztng/tt-bio/releases/tag/v0.6.2", "title": "tt-bio v0.6.2", "content_html": "OpenFold3 is now live as tt-bio predict --model openfold3, bringing the OpenFold Consortium's AlphaFold3 reproduction to Tenstorrent hardware—it folds proteins, RNA, and DNA chains with per-chain MSAs and optional templates, reusing the same scheduler, multi-card fan-out, and MSA cache as Protenix-v2. The model is gated against the official CPU reference across seven accuracy and parity legs, runs the on-device diffusion sampler in fp32 by default for parity, and strictly handles unsupported features (ligands, covalent bonds, PAE output) by raising rather than silently degrading; note that you'll need to fetch and point the OF3_CKPT environment variable at the consortium's public weights yourself. A handful of fixes shore up the debut—missing dependency declarations that broke clean installs, progress reporting wired through the trunk and sampler loops, and a tightened chunk gate (W<=384) that closed a 512-residue memory overflow—plus 1.18× GPU scaling improvements in the 300-residue range landed across Protenix-v2 and OpenDDE through kernel re-tuning (parity docs).
Links: Repository · v0.6.2
\naffiliated · added 2026-08-07
", "summary": "OpenFold3 is now live as `tt-bio predict --model openfold3`, bringing the OpenFold Consortium's AlphaFold3 reproduction to Tenstorrent hardware—it folds proteins, RNA, and DNA chains with per-chain MSAs and optional templates, reusing the same scheduler, multi-card fan-out, and MSA cache as Protenix-v2. The model is gated against the official CPU reference across seven accuracy and parity legs, runs the on-device diffusion sampler in fp32 by default for parity, and strictly handles unsupported features (ligands, covalent bonds, PAE output) by raising rather than silently degrading; note that you'll need to fetch and point the `OF3_CKPT` environment variable at the consortium's public weights yourself. A handful of fixes shore up the debut—missing dependency declarations that broke clean installs, progress reporting wired through the trunk and sampler loops, and a tightened chunk gate (`W<=384`) that closed a 512-residue memory overflow—plus 1.18× GPU scaling improvements in the 300-residue range landed across Protenix-v2 and OpenDDE through kernel re-tuning ([parity docs](https://github.com/moritztng/tt-bio/blob/main/docs/implementation-parity.md)).", "date_published": "2026-08-07T21:33:46Z", "tags": [ "affiliated", "release" ] }, { "id": "https://github.com/tenstorrent/luwen/releases/tag/bh-mod-v2.0.0", "url": "https://github.com/tenstorrent/luwen/releases/tag/bh-mod-v2.0.0", "title": "luwen bh-mod-v2.0.0", "content_html": "This release hardens DRAM harvest configuration on Blackhole by validating harvest field values before writing to the ccfgovr register and rejecting out-of-range values that the chip cannot accept—preventing silent misconfiguration that could lead to memory instability or undefined behavior downstream. If you're tuning Blackhole memory parameters, these validation checks ensure your settings actually match hardware constraints.
Links: Repository · bh-mod-v2.0.0
\nofficial · added 2026-08-07
", "summary": "This release hardens DRAM harvest configuration on Blackhole by validating harvest field values before writing to the `ccfgovr` register and rejecting out-of-range values that the chip cannot accept—preventing silent misconfiguration that could lead to memory instability or undefined behavior downstream. If you're tuning Blackhole memory parameters, these [validation checks](https://github.com/tenstorrent/luwen/pull/167) ensure your settings actually match hardware constraints.", "date_published": "2026-08-07T14:05:56Z", "tags": [ "official", "release" ] }, { "id": "https://github.com/tenstorrent/sfpi/releases/tag/7.68.0", "url": "https://github.com/tenstorrent/sfpi/releases/tag/7.68.0", "title": "SFPI 7.68.0", "content_html": "The SFPI compiler now supports additional builtin combination operations—though some require the -ffast-math flag to unlock—giving you more flexibility in fusing operations on-device. We've also fixed a subtle bug in ttreplay optimization that could leave the replay buffer partially empty, which should improve both performance and predictability for workloads relying on replay. A handful of stability and efficiency improvements round out the release, along with deprecation of vBool-to-vInt conversion, which had no real-world uses anyway.
Links: Repository · 7.68.0
\nofficial · added 2026-08-07
", "summary": "The SFPI compiler now supports additional builtin combination operations—though some require the `-ffast-math` flag to unlock—giving you more flexibility in fusing operations on-device. We've also fixed a subtle bug in `ttreplay` optimization that could leave the replay buffer partially empty, which should improve both performance and predictability for workloads relying on replay. A handful of stability and efficiency improvements round out the release, along with deprecation of `vBool`-to-`vInt` conversion, which had no real-world uses anyway.", "date_published": "2026-08-07T13:57:54Z", "tags": [ "official", "release" ] }, { "id": "https://github.com/dstackai/dstack/releases/tag/0.21.0", "url": "https://github.com/dstackai/dstack/releases/tag/0.21.0", "title": "dstack 0.21.0", "content_html": "The jump to Pydantic v2 unblocks Python 3.14 support and opens the door to newer ecosystem libraries, while bringing faster validation and CLI startup — but if you're using the Python API or have dstack plugins, you'll need to verify they're compatible with Pydantic v2 models before upgrading. Beyond the foundational work, this release brings several UX wins: dstack metrics now shows CPU, memory, and GPU utilization as sparklines over the last hour so you can spot trends at a glance; AWS gateways with ACM certificates can now run multiple replicas with HTTPS handled directly by dstack instead of requiring an external load balancer; and the preset system gained baseline trials, per-trial learnings, and a richer CLI view to make benchmark results more interpretable. You'll also see GPU driver detection in dstack fleet -v and new --full-offers and --unallocated flags for better capacity discovery on Kubernetes and Slurm. Note that servers and CLIs must be upgraded together — new CLIs won't work with 0.20.x servers.
Links: Repository · 0.21.0
\ncommunity · added 2026-08-06
", "summary": "The jump to Pydantic v2 unblocks Python 3.14 support and opens the door to newer ecosystem libraries, while bringing faster validation and CLI startup — but if you're using the Python API or have `dstack` plugins, you'll need to verify they're compatible with Pydantic v2 models before upgrading. Beyond the foundational work, this release brings several UX wins: `dstack metrics` now shows CPU, memory, and GPU utilization as sparklines over the last hour so you can spot trends at a glance; AWS gateways with ACM certificates can now run multiple replicas with HTTPS handled directly by `dstack` instead of requiring an external load balancer; and the preset system gained [baseline trials, per-trial learnings, and a richer CLI view](https://github.com/dstackai/dstack/pull/4078) to make benchmark results more interpretable. You'll also see GPU driver detection in `dstack fleet -v` and new `--full-offers` and `--unallocated` flags for better capacity discovery on Kubernetes and Slurm. Note that servers and CLIs must be upgraded together — new CLIs won't work with 0.20.x servers.", "date_published": "2026-08-06T13:31:14Z", "tags": [ "community", "release" ] }, { "id": "https://github.com/tenstorrent/tt-rpm", "url": "https://github.com/tenstorrent/tt-rpm", "title": "tt-rpm", "content_html": "Cycle-level, execution-driven RISC-V CPU performance model built on Sparta (MAP) with Whisper supplying functional execution, so it runs real ELF binaries — CoreMark, Dhrystone — to completion. The pipeline is YAML-configurable across in-order/out-of-order execution, issue policy, execute granularity, write-port arbitration, and bypass paths, with a modeled L1 I$/D$ plus optional unified L2, per-unit logging, stats reports, and Konata pipeline visualization.
\n\nLinks: Repo
\nofficial · added 2026-08-05
\nrisc-v, performance-model, simulation, sparta, whisper, microarchitecture, riscv-arch, dev-tools
", "summary": "Cycle-level, execution-driven RISC-V CPU performance model built on Sparta (MAP) with Whisper supplying functional execution, so it runs real ELF binaries — CoreMark, Dhrystone — to completion. The pipeline is YAML-configurable across in-order/out-of-order execution, issue policy, execute granularity, write-port arbitration, and bypass paths, with a modeled L1 I$/D$ plus optional unified L2, per-unit logging, stats reports, and Konata pipeline visualization.", "date_published": "2026-08-05T23:59:51Z", "tags": [ "riscv-arch", "dev-tools", "official", "entry" ] }, { "id": "https://github.com/tenstorrent/vllm-tt-plugin", "url": "https://github.com/tenstorrent/vllm-tt-plugin", "title": "vllm-tt-plugin", "content_html": "Tenstorrent backend for vLLM, built on vLLM's standard plugin mechanism — install it alongside vLLM and TT hardware registers itself as a platform whenever ttnn is importable. Self-contained: model registration, platform detection, scheduling, worker execution, model loading, async decode, and data-parallel/multi-lane execution all live in the plugin, so nothing Tenstorrent-specific has to land in vLLM core.
Links: Repo
\nofficial · added 2026-08-05
\nvllm, serving, inference, llm, openai-compatible, plugin, ai-models, cloud-infra
", "summary": "Tenstorrent backend for vLLM, built on vLLM's standard plugin mechanism — install it alongside vLLM and TT hardware registers itself as a platform whenever `ttnn` is importable. Self-contained: model registration, platform detection, scheduling, worker execution, model loading, async decode, and data-parallel/multi-lane execution all live in the plugin, so nothing Tenstorrent-specific has to land in vLLM core.", "date_published": "2026-08-05T23:59:50Z", "tags": [ "ai-models", "cloud-infra", "official", "entry" ] }, { "id": "https://github.com/tenstorrent/tt-low-level-documentation", "url": "https://github.com/tenstorrent/tt-low-level-documentation", "title": "tt-low-level-documentation", "content_html": "Documentation for the low-level layer of tt-metal: compute LLK APIs and data movement APIs. The data movement side covers the NOC and overlay on Wormhole and Blackhole; the compute side covers Tensix hardware and expected usage of the LLK APIs. Aimed at op and model writers who need to know what the APIs do and how the hardware behaves underneath them.
\n\nLinks: Repo
\nofficial · added 2026-08-05
\nllk, data-movement, noc, tensix, tt-metal, documentation, guides, kernels
", "summary": "Documentation for the low-level layer of tt-metal: compute LLK APIs and data movement APIs. The data movement side covers the NOC and overlay on Wormhole and Blackhole; the compute side covers Tensix hardware and expected usage of the LLK APIs. Aimed at op and model writers who need to know what the APIs do and how the hardware behaves underneath them.", "date_published": "2026-08-05T23:59:49Z", "tags": [ "guides", "kernels", "official", "entry" ] }, { "id": "https://github.com/tenstorrent/tt-CableGen", "url": "https://github.com/tenstorrent/tt-CableGen", "title": "tt-CableGen", "content_html": "Network cabling visualizer for Tenstorrent scale-out deployments: describe a target topology and it generates and renders how to physically cable multiple Wormhole or Blackhole systems together. Works in a physical-deployment mode with racking information and a logical-hierarchy mode for clustering/pod groupings, with topology import/export and a Docker deployment path.
\n\nLinks: Repo
\nofficial · added 2026-08-05
\ncabling, scale-out, topology, multi-host, visualization, hw-system, cloud-infra
", "summary": "Network cabling visualizer for Tenstorrent scale-out deployments: describe a target topology and it generates and renders how to physically cable multiple Wormhole or Blackhole systems together. Works in a physical-deployment mode with racking information and a logical-hierarchy mode for clustering/pod groupings, with topology import/export and a Docker deployment path.", "date_published": "2026-08-05T23:59:48Z", "tags": [ "hw-system", "cloud-infra", "official", "entry" ] }, { "id": "https://github.com/tenstorrent/tt-kernel-package-manager", "url": "https://github.com/tenstorrent/tt-kernel-package-manager", "title": "tt-kernel", "content_html": "Distributes models over the Hugging Face Hub and serves them on Tenstorrent hardware — tt-kernel serve <namespace>/<model> pulls a bundle, registers it with the Tenstorrent vLLM plugin, and launches an OpenAI-compatible server. A vLLM bundle ships only adapter code and metadata (weights stay referenced by HF repo id), while legacy kernel-cache bundles package precompiled tt-metal kernel binaries so a model's first run is a cache hit instead of a slow JIT recompile. Explicitly experimental — the bundle format and APIs may change without notice.
Links: Repo
\nofficial · added 2026-08-05
\nkernel-cache, huggingface, package-manager, vllm, tt-metal, serving, dev-tools, kernels, ai-models
", "summary": "Distributes models over the Hugging Face Hub and serves them on Tenstorrent hardware — `tt-kernel serveSingle entry point to the Tenstorrent software stack: tt update converges a machine onto the CI-tested "golden" version set, tt device covers status/info/reset, and tt model/tt serve pull weights and bring up tt-inference-server. Commands either run natively or delegate to tt-smi, tt-flash, and tt-installer behind a stable interface, with --json output and documented exit codes on every command. Early prototype — the README labels it an internal prototype and several subcommands are still stubs.
Links: Repo
\nofficial · added 2026-08-05
\ncli, device-management, model-serving, tt-smi, prototype, dev-tools, hw-system
", "summary": "Single entry point to the Tenstorrent software stack: `tt update` converges a machine onto the CI-tested \"golden\" version set, `tt device` covers status/info/reset, and `tt model`/`tt serve` pull weights and bring up tt-inference-server. Commands either run natively or delegate to tt-smi, tt-flash, and tt-installer behind a stable interface, with `--json` output and documented exit codes on every command. Early prototype — the README labels it an internal prototype and several subcommands are still stubs.", "date_published": "2026-08-05T23:59:46Z", "tags": [ "dev-tools", "hw-system", "official", "entry" ] }, { "id": "https://github.com/tenstorrent/ttnn-visualizer/releases/tag/v0.97.0", "url": "https://github.com/tenstorrent/ttnn-visualizer/releases/tag/v0.97.0", "title": "ttnn-visualizer v0.97.0", "content_html": "The visualizer now renders cluster topology and placement directly from the cluster descriptor without requiring a baked SoC descriptor, making it easier to inspect multi-chip systems on the fly—and when you do have per-chip arch data, it's applied as optional enrichment rather than a hard blocker. Multi-host profiling is cleaner too: remote sync can now discover and download performance reports per rank, and reads default to rank 0 to eliminate the visual collision of overlapping data from every rank. A few quality-of-life fixes round out the release—version checks no longer flag patch updates, SSH config is read to auto-populate remote connections, and performance charts got a tidier layout with grouped headings.
\n\nLinks: Repository · v0.97.0
\nofficial · added 2026-08-05
", "summary": "The visualizer now renders cluster topology and placement directly from the cluster descriptor without requiring a baked SoC descriptor, making it easier to inspect multi-chip systems on the fly—and when you do have per-chip arch data, it's applied as optional enrichment rather than a hard blocker. Multi-host profiling is cleaner too: remote sync can now discover and download performance reports per rank, and reads default to rank 0 to eliminate the visual collision of overlapping data from every rank. A few quality-of-life fixes round out the release—[version checks no longer flag patch updates](https://github.com/tenstorrent/ttnn-visualizer/pull/1820), SSH config is read to auto-populate remote connections, and performance charts got a tidier layout with grouped headings.", "date_published": "2026-08-05T17:46:17Z", "tags": [ "official", "release" ] }, { "id": "https://github.com/tsingletaryTT/tensix-viz/releases/tag/v1.2.0", "url": "https://github.com/tsingletaryTT/tensix-viz/releases/tag/v1.2.0", "title": "tensix-viz v1.2.0", "content_html": "tensix-viz v1.2.0 sharpens memory stats configuration for hardware-powered visuals, making it easier to instrument and monitor Tensix performance directly from visualization tools. This matters when you're profiling real workloads on silicon and need reliable memory telemetry baked into your dashboards without extra plumbing. Check the full changelog for implementation details.
\n\nLinks: Repository · v1.2.0
\naffiliated · added 2026-08-05
", "summary": "tensix-viz v1.2.0 sharpens memory stats configuration for hardware-powered visuals, making it easier to instrument and monitor Tensix performance directly from visualization tools. This matters when you're profiling real workloads on silicon and need reliable memory telemetry baked into your dashboards without extra plumbing. Check the [full changelog](https://github.com/tsingletaryTT/tensix-viz/compare/v1.1.2...v1.2.0) for implementation details.", "date_published": "2026-08-05T16:35:19Z", "tags": [ "affiliated", "release" ] }, { "id": "https://github.com/tenstorrent/tt-installer/releases/tag/v3.5.2", "url": "https://github.com/tenstorrent/tt-installer/releases/tag/v3.5.2", "title": "tt-installer v3.5.2", "content_html": "UV integration now respects existing virtual environments, so you won't hit friction if you're working in an already-active venv during installation—a small but meaningful quality-of-life improvement for developers with existing Python setups.
\n\nLinks: Repository · v3.5.2
\nofficial · added 2026-07-30
", "summary": "UV integration now respects existing virtual environments, so you won't hit friction if you're working in an already-active venv during installation—a small but meaningful quality-of-life improvement for developers with existing Python setups.", "date_published": "2026-07-30T18:00:15Z", "tags": [ "official", "release" ] }, { "id": "https://github.com/tenstorrent/tt-forge/releases/tag/1.4.0", "url": "https://github.com/tenstorrent/tt-forge/releases/tag/1.4.0", "title": "tt-forge 1.4.0", "content_html": "tt-forge 1.4.0 brings a comprehensive refresh of model coverage and infrastructure, now supporting everything from Llama 3.2 Vision to Qwen 3 and Falcon 3 across both n150 and p150 hardware. The most significant change is the expanded LLM lineup—you can now compile and deploy over 40 language model variants including the latest reasoning-focused OLMo3 models and multimodal systems like Pixtral—making it practical to benchmark performance across a much wider range of production targets. On the non-LLM side, embedding models, vision transformers, and diffusion backends round out the supported model matrix, while internal tooling improvements (including a move of AI bringup scripts to a dedicated repo) clean up the monorepo and streamline testing workflows. The release includes updated PyTorch and dependency commits across tt-xla, tt-mlir, and tt-metal, ensuring compatibility with recent compiler and hardware stacks.
\n\nLinks: Repository · 1.4.0
\nofficial · added 2026-07-30
", "summary": "tt-forge 1.4.0 brings a comprehensive refresh of model coverage and infrastructure, now supporting everything from Llama 3.2 Vision to Qwen 3 and Falcon 3 across both n150 and p150 hardware. The most significant change is the expanded LLM lineup—you can now compile and deploy over 40 language model variants including the latest reasoning-focused OLMo3 models and multimodal systems like Pixtral—making it practical to benchmark performance across a much wider range of production targets. On the non-LLM side, embedding models, vision transformers, and diffusion backends round out the [supported model matrix](https://github.com/tenstorrent/tt-forge-models), while internal tooling improvements (including a move of AI bringup scripts to a dedicated repo) clean up the monorepo and streamline testing workflows. The release includes updated PyTorch and dependency commits across tt-xla, tt-mlir, and tt-metal, ensuring compatibility with recent compiler and hardware stacks.", "date_published": "2026-07-30T11:17:16Z", "tags": [ "official", "release" ] }, { "id": "https://github.com/tenstorrent/tt-forge-onnx/releases/tag/1.4.0", "url": "https://github.com/tenstorrent/tt-forge-onnx/releases/tag/1.4.0", "title": "tt-forge-onnx 1.4.0", "content_html": "This release rolls up steady progress across tt-mlir and tt-forge-models dependencies, keeping the compiler stack current with upstream development. The highlight is a documentation migration from mdBook to Sphinx, which gives you a refreshed UI and improved navigation for guides and API references—worth checking out if you've been following along with the docs.
\n\nLinks: Repository · 1.4.0
\nofficial · added 2026-07-30
", "summary": "This release rolls up steady progress across tt-mlir and tt-forge-models dependencies, keeping the compiler stack current with upstream development. The highlight is a [documentation migration from mdBook to Sphinx](https://github.com/tenstorrent/tt-forge-onnx/pull/3381), which gives you a refreshed UI and improved navigation for guides and API references—worth checking out if you've been following along with the docs.", "date_published": "2026-07-30T11:12:02Z", "tags": [ "official", "release" ] }, { "id": "https://github.com/tenstorrent/tt-xla/releases/tag/1.4.0", "url": "https://github.com/tenstorrent/tt-xla/releases/tag/1.4.0", "title": "tt-xla 1.4.0", "content_html": "Chunked prefill with decoupled buckets and runtime scheduling, plus multi-parallelism support across generative and pooling models, unlock more efficient vLLM inference on Tenstorrent hardware—letting you scale from single-device to data-parallel and tensor-parallel deployments without manual graph fragmentation. Beyond vLLM, you'll find new e2e generative model pipelines (Flux, Flux.2, Janus-Pro, Z-Image, GLM-Image), MLA attention in vLLM, mixed-precision quantization heuristics, improved SDPA handling for bool masks behind an opt-in flag, and fixes to device-count mismatches and layer-norm dtype issues under AOTAutograd, all tested against a growing roster of LLMs and image-generation models across n150, p150, and n300 hardware.
\n\nLinks: Repository · 1.4.0
\nofficial · added 2026-07-30
", "summary": "Chunked prefill with decoupled buckets and runtime scheduling, plus multi-parallelism support across generative and pooling models, unlock more efficient vLLM inference on Tenstorrent hardware—letting you scale from single-device to data-parallel and tensor-parallel deployments without manual graph fragmentation. Beyond vLLM, you'll find new e2e generative model pipelines (Flux, Flux.2, Janus-Pro, Z-Image, GLM-Image), MLA attention in vLLM, mixed-precision quantization heuristics, improved SDPA handling for bool masks behind an opt-in flag, and fixes to device-count mismatches and layer-norm dtype issues under AOTAutograd, all tested against a growing roster of LLMs and image-generation models across n150, p150, and n300 hardware.", "date_published": "2026-07-30T11:06:22Z", "tags": [ "official", "release" ] }, { "id": "https://github.com/tenstorrent/tt-studio/releases/tag/rc-v2.9.0", "url": "https://github.com/tenstorrent/tt-studio/releases/tag/rc-v2.9.0", "title": "TT-Studio rc-v2.9.0", "content_html": "Studio's first-run experience moves into the app itself with in-app setup and secrets management, eliminating the .env friction for new users, and you can now edit the Voice Agent's system prompt at runtime to tune behavior on the fly. The workflow canvas got a substantial refresh—model auto-select, live code streaming, dynamic edge styling with smart deletion guards, and auto-picked RAG collections—alongside fresh model support (Flux Schnell) and parallel deployment to speed up initialization. A backend-unreachable overlay provides clearer feedback when connectivity drops, and the launcher now handles branch-switching and cleanup via --uninstall and --switch flags. The backend bumps to v0.18.0 of tt-inference-server and consolidates state management to keep the source tree clean.
Links: Repository · rc-v2.9.0
\nofficial · added 2026-07-28
", "summary": "Studio's first-run experience moves into the app itself with in-app setup and secrets management, eliminating the `.env` friction for new users, and you can now edit the Voice Agent's system prompt at runtime to tune behavior on the fly. The workflow canvas got a substantial refresh—model auto-select, live code streaming, dynamic edge styling with smart deletion guards, and auto-picked RAG collections—alongside fresh model support (Flux Schnell) and parallel deployment to speed up initialization. A [backend-unreachable overlay](https://github.com/tenstorrent/tt-studio/issues/1157) provides clearer feedback when connectivity drops, and the launcher now handles branch-switching and cleanup via `--uninstall` and `--switch` flags. The backend bumps to `v0.18.0` of `tt-inference-server` and consolidates state management to keep the source tree clean.", "date_published": "2026-07-28T18:09:27Z", "tags": [ "official", "release" ] }, { "id": "https://github.com/tenstorrent/tt-inference-server/releases/tag/v0.19.0", "url": "https://github.com/tenstorrent/tt-inference-server/releases/tag/v0.19.0", "title": "tt-inference-server v0.19.0", "content_html": "Llama-3.1-8B-Instruct is now running on P300 hardware, expanding the inference server's device support for this widely-used instruction-tuned model. The uplift tracks the latest TT-Metal commits and maintains functional status, letting you deploy this capable 8B model on newer Tenstorrent accelerators.
\n\nLinks: Repository · v0.19.0
\nofficial · added 2026-07-24
", "summary": "Llama-3.1-8B-Instruct is now running on P300 hardware, expanding the inference server's device support for this widely-used instruction-tuned model. The uplift tracks the latest TT-Metal commits and maintains functional status, letting you deploy this capable 8B model on newer Tenstorrent accelerators.", "date_published": "2026-07-24T16:58:16Z", "tags": [ "official", "release" ] }, { "id": "https://github.com/tenstorrent/tt-topology/releases/tag/v1.2.20", "url": "https://github.com/tenstorrent/tt-topology/releases/tag/v1.2.20", "title": "tt-topology v1.2.20", "content_html": "This release bumps the underlying pyluwen dependency to 0.8.5, bringing in improvements and fixes from that library that support tt-topology's hardware introspection and topology validation work. If you're using tt-topology to characterize or configure your Tenstorrent mesh, this update ensures you're working with the latest device communication layer.
Links: Repository · v1.2.20
\nofficial · added 2026-07-23
", "summary": "This release bumps the underlying `pyluwen` dependency to 0.8.5, bringing in improvements and fixes from that library that support tt-topology's hardware introspection and topology validation work. If you're using tt-topology to characterize or configure your Tenstorrent mesh, this update ensures you're working with the latest device communication layer.", "date_published": "2026-07-23T15:35:12Z", "tags": [ "official", "release" ] }, { "id": "https://github.com/tenstorrent/tt-exalens/releases/tag/v0.3.29", "url": "https://github.com/tenstorrent/tt-exalens/releases/tag/v0.3.29", "title": "tt-exalens v0.3.29", "content_html": "Exalens now supports step functionality for Rocket cores, letting you pause and advance through Rocket code one instruction at a time—a critical addition for debugging complex kernel behavior. The release also improves memory introspection by exposing read and write memory bytes in the debug interface, and fixes a longstanding issue where noc_read32 and noc_write32 were incorrectly routing through DMA instead of executing directly, which should resolve performance anomalies in code relying on those primitives.
Links: Repository · v0.3.29
\nofficial · added 2026-07-22
", "summary": "Exalens now supports step functionality for Rocket cores, letting you pause and advance through Rocket code one instruction at a time—a critical addition for debugging complex kernel behavior. The release also improves memory introspection by exposing read and write memory bytes in the debug interface, and fixes a longstanding issue where `noc_read32` and `noc_write32` were incorrectly routing through DMA instead of executing directly, which should resolve performance anomalies in code relying on those primitives.", "date_published": "2026-07-22T10:37:50Z", "tags": [ "official", "release" ] }, { "id": "https://github.com/moritztng/tt-atom/releases/tag/v0.2.1", "url": "https://github.com/moritztng/tt-atom/releases/tag/v0.2.1", "title": "tt-atom v0.2.1", "content_html": "The unified entry point Calculator(atoms, model=...) cuts through three separate factory functions, dispatching automatically by checkpoint name just like FAIRChemCalculator or HF's AutoModel—a clean API win that makes Orb-v3 and UMA feel like one family to users. Beyond that surface change, OrbCalculator.evaluate_batch delivers measured 12–19× speedup on Orb inference (outpacing UMA's own ~13× batched gain), and the release consolidates the whole pipeline: fused_lnbw now defaults on when the hardware supports it, a machine-readable release gate replaces manual checklist steps, and the MultiCard worker—which was silently building only the UMA path—now correctly dispatches to Orb. Honest perf-per-dollar numbers land too: the p150 trails the H200 on raw throughput but wins by ~4–9× per dollar as systems scale, verified with committed raw timings; an earlier 40× claim that lacked evidence is withdrawn. Several measured fusion wins (closed-form geometry VJP, fused SiLU backward, row-major scatter) compound to real speedups, and a new MgO rock-salt multi-element bulk baseline rounds out the parity table.
Links: Repository · v0.2.1
\naffiliated · added 2026-07-20
", "summary": "The unified entry point `Calculator(atoms, model=...)` cuts through three separate factory functions, dispatching automatically by checkpoint name just like `FAIRChemCalculator` or HF's `AutoModel`—a clean API win that makes Orb-v3 and UMA feel like one family to users. Beyond that surface change, `OrbCalculator.evaluate_batch` delivers measured 12–19× speedup on Orb inference (outpacing UMA's own ~13× batched gain), and the release consolidates the whole pipeline: `fused_lnbw` now defaults on when the hardware supports it, a [machine-readable release gate](https://github.com/moritztng/tt-atom/blob/main/scripts/release_gate.py) replaces manual checklist steps, and the MultiCard worker—which was silently building only the UMA path—[now correctly dispatches to Orb](https://github.com/moritztng/tt-atom/blob/main/tt_atom/batch.py). Honest perf-per-dollar numbers land too: the p150 trails the H200 on raw throughput but wins by ~4–9× per dollar as systems scale, verified with committed raw timings; an earlier 40× claim that lacked evidence is withdrawn. Several measured fusion wins (closed-form geometry VJP, fused SiLU backward, row-major scatter) compound to real speedups, and [a new MgO rock-salt multi-element bulk baseline](https://github.com/moritztng/tt-atom/blob/main/docs/materials-benchmark.md) rounds out the parity table.", "date_published": "2026-07-20T11:06:13Z", "tags": [ "affiliated", "release" ] }, { "id": "https://github.com/tenstorrent/sfpi", "url": "https://github.com/tenstorrent/sfpi", "title": "SFPI", "content_html": "Tenstorrent SFPU programming interface — TT-enhanced RISC-V GCC and binutils plus header files for programming the Tensix SFPU (vector engine) from kernel code. The compiler toolchain underneath TT-Metalium's SFPU ops.
\n\nLinks: Repo
\nofficial · added 2026-07-17
\nsfpu, compiler-toolchain, gcc, riscv, kernels, dev-tools
", "summary": "Tenstorrent SFPU programming interface — TT-enhanced RISC-V GCC and binutils plus header files for programming the Tensix SFPU (vector engine) from kernel code. The compiler toolchain underneath TT-Metalium's SFPU ops.", "date_published": "2026-07-17T23:59:45Z", "tags": [ "kernels", "dev-tools", "official", "entry" ] }, { "id": "https://github.com/tenstorrent/tt-tools-common", "url": "https://github.com/tenstorrent/tt-tools-common", "title": "tt-tools-common", "content_html": "Shared helper library of common utilities used across Tenstorrent system tools such as tt-smi, tt-flash, and tt-topology. A dependency rather than a standalone tool.
\n\nLinks: Repo
\nofficial · added 2026-07-17
\nlibrary, tooling, shared-utilities, hw-system
", "summary": "Shared helper library of common utilities used across Tenstorrent system tools such as tt-smi, tt-flash, and tt-topology. A dependency rather than a standalone tool.", "date_published": "2026-07-17T23:59:44Z", "tags": [ "hw-system", "official", "entry" ] }, { "id": "https://github.com/tenstorrent/tt-system-tools", "url": "https://github.com/tenstorrent/tt-system-tools", "title": "tt-system-tools", "content_html": "System setup and support utilities for Tenstorrent hardware — hugepages-setup configures the 1GB hugepages TT ASICs need, and tt-oops collects diagnostic data for troubleshooting. Ships as the tenstorrent-tools deb/rpm.
\n\nLinks: Repo
\nofficial · added 2026-07-17
\nhugepages, system-setup, diagnostics, hw-system
", "summary": "System setup and support utilities for Tenstorrent hardware — hugepages-setup configures the 1GB hugepages TT ASICs need, and tt-oops collects diagnostic data for troubleshooting. Ships as the tenstorrent-tools deb/rpm.", "date_published": "2026-07-17T23:59:43Z", "tags": [ "hw-system", "official", "entry" ] }, { "id": "https://github.com/tenstorrent/tt-burnin", "url": "https://github.com/tenstorrent/tt-burnin", "title": "tt-burnin", "content_html": "Command-line utility that runs a high power-consumption workload on Tenstorrent devices — used for chip testing, burn-in, and validating a system's power delivery and cooling under sustained load.
\n\nLinks: Repo
\nofficial · added 2026-07-17
\nburn-in, stress-test, power, hardware-validation, hw-system
", "summary": "Command-line utility that runs a high power-consumption workload on Tenstorrent devices — used for chip testing, burn-in, and validating a system's power delivery and cooling under sustained load.", "date_published": "2026-07-17T23:59:42Z", "tags": [ "hw-system", "official", "entry" ] }, { "id": "https://github.com/tenstorrent/tt-vscode-toolkit/releases/tag/v0.1.19", "url": "https://github.com/tenstorrent/tt-vscode-toolkit/releases/tag/v0.1.19", "title": "tt-vscode-toolkit v0.1.19", "content_html": "The VS Code toolkit now includes a comprehensive guide to joyfully monkeypatching TT-NN without modifying your precious tt-metal checkouts.
\n\nLinks: Repository · v0.1.19
\nofficial · added 2026-07-16
", "summary": "The VS Code toolkit now includes a comprehensive guide to joyfully monkeypatching TT-NN without modifying your precious tt-metal checkouts.", "date_published": "2026-07-16T21:17:06Z", "tags": [ "official", "release" ] }, { "id": "https://github.com/tenstorrent/tt-burnin/releases/tag/v0.4.3", "url": "https://github.com/tenstorrent/tt-burnin/releases/tag/v0.4.3", "title": "tt-burnin v0.4.3", "content_html": "Galaxy 6U systems get a safer stress-test exit: the tool now performs the KMD reset handshake when resetting Blackhole Galaxy hardware (#37), so the kernel driver and the burn-in workload agree about device state instead of leaving a rack-scale box in limbo after a run.
\n\nLinks: Repository · v0.4.3
\nofficial · added 2026-07-16
", "summary": "Galaxy 6U systems get a safer stress-test exit: the tool now performs the KMD reset handshake when resetting Blackhole Galaxy hardware ([#37](https://github.com/tenstorrent/tt-burnin/issues/37)), so the kernel driver and the burn-in workload agree about device state instead of leaving a rack-scale box in limbo after a run.", "date_published": "2026-07-16T15:26:18Z", "tags": [ "official", "release" ] }, { "id": "https://github.com/tenstorrent/tt-kmd/releases/tag/ttkmd-2.10.0", "url": "https://github.com/tenstorrent/tt-kmd/releases/tag/ttkmd-2.10.0", "title": "tt-kmd ttkmd-2.10.0", "content_html": "The driver now allows exclusive device access via O_EXCL flags, letting tools like tt-flash guarantee they hold the device for the entire flash-and-reset cycle without losing ownership across intermediate resets. A new experimental ioctl exports TLB windows as dma-bufs for peer-to-peer PCIe DMA from devices like RDMA NICs directly into chip memory—though the release notes carefully document the hazards when a NOC hangs mid-DMA, which can escalate to host machine-check exceptions. The update also fixes resource leaks where wedged processes could exhaust TLB and iATU pools across resets, and corrects a 58-bit address masking bug on Blackhole that was silently misdirecting pinned-buffer DMA to the wrong outbound window.
Links: Repository · ttkmd-2.10.0
\nofficial · added 2026-07-13
", "summary": "The driver now allows exclusive device access via `O_EXCL` flags, letting tools like `tt-flash` guarantee they hold the device for the entire flash-and-reset cycle without losing ownership across intermediate resets. A new experimental ioctl exports TLB windows as dma-bufs for peer-to-peer PCIe DMA from devices like RDMA NICs directly into chip memory—though the release notes carefully document the hazards when a NOC hangs mid-DMA, which can escalate to host machine-check exceptions. The update also fixes resource leaks where wedged processes could exhaust TLB and iATU pools across resets, and corrects a 58-bit address masking bug on Blackhole that was silently misdirecting pinned-buffer DMA to the wrong outbound window.", "date_published": "2026-07-13T20:07:36Z", "tags": [ "official", "release" ] }, { "id": "https://github.com/tenstorrent/riscv_arch_tests", "url": "https://github.com/tenstorrent/riscv_arch_tests", "title": "riscv_arch_tests", "content_html": "RISC-V architectural self-checking directed tests — randomly-generated register operands and data with low-level OS code for test scheduling and self-checking, runnable on a RISC-V design or an ISS such as Whisper or Spike. Generated by an internal Tenstorrent tool from the official RISC-V ISA spec.
\n\nLinks: Repo
\nofficial · added 2026-07-10
\nriscv, testing, verification, isa, architecture, riscv-arch
", "summary": "RISC-V architectural self-checking directed tests — randomly-generated register operands and data with low-level OS code for test scheduling and self-checking, runnable on a RISC-V design or an ISS such as Whisper or Spike. Generated by an internal Tenstorrent tool from the official RISC-V ISA spec.", "date_published": "2026-07-10T23:59:41Z", "tags": [ "riscv-arch", "official", "entry" ] }, { "id": "https://github.com/tenstorrent/tt-isa-documentation", "url": "https://github.com/tenstorrent/tt-isa-documentation", "title": "tt-isa-documentation", "content_html": "Low-level ISA and microarchitecture documentation for Tenstorrent AI architectures (Grayskull, Wormhole, Blackhole) — the authoritative hardware reference beneath the tt-forge / tt-metal software stack.
\n\nLinks: Repo
\nofficial · added 2026-07-10
\nisa, architecture, documentation, tensix, low-level, research, guides
", "summary": "Low-level ISA and microarchitecture documentation for Tenstorrent AI architectures (Grayskull, Wormhole, Blackhole) — the authoritative hardware reference beneath the tt-forge / tt-metal software stack.", "date_published": "2026-07-10T23:59:40Z", "tags": [ "research", "guides", "official", "entry" ] }, { "id": "https://github.com/tenstorrent/polaris", "url": "https://github.com/tenstorrent/polaris", "title": "polaris", "content_html": "A high-level AI simulator from Tenstorrent for modeling and exploring AI accelerator and workload performance.
\n\nLinks: Repo
\nofficial · added 2026-07-10
\nsimulator, performance, modeling, architecture, research, dev-tools
", "summary": "A high-level AI simulator from Tenstorrent for modeling and exploring AI accelerator and workload performance.", "date_published": "2026-07-10T23:59:39Z", "tags": [ "research", "dev-tools", "official", "entry" ] }, { "id": "https://github.com/tenstorrent/tt-forge-models", "url": "https://github.com/tenstorrent/tt-forge-models", "title": "tt-forge-models", "content_html": "A shared repository of model implementations used across TT-Forge frontends — a single source of truth for the models used in testing and benchmarking, rather than duplicating them across frontend repos.
\n\nLinks: Repo
\nofficial · added 2026-07-10
\ntt-forge, models, benchmarking, testing, inference, ai-models
", "summary": "A shared repository of model implementations used across TT-Forge frontends — a single source of truth for the models used in testing and benchmarking, rather than duplicating them across frontend repos.", "date_published": "2026-07-10T23:59:38Z", "tags": [ "ai-models", "official", "entry" ] }, { "id": "https://github.com/tenstorrent/tt-perf-report", "url": "https://github.com/tenstorrent/tt-perf-report", "title": "tt-perf-report", "content_html": "Performance report analysis tool for Tenstorrent Metal operations — analyzes perf traces to surface throughput, bottlenecks, and optimization opportunities.
\n\nLinks: Repo
\nofficial · added 2026-07-10
\nperformance, profiling, tt-metal, analysis, optimization, dev-tools
", "summary": "Performance report analysis tool for Tenstorrent Metal operations — analyzes perf traces to surface throughput, bottlenecks, and optimization opportunities.", "date_published": "2026-07-10T23:59:37Z", "tags": [ "dev-tools", "official", "entry" ] }, { "id": "https://github.com/tenstorrent/tt-emule", "url": "https://github.com/tenstorrent/tt-emule", "title": "tt-emule", "content_html": "A C++ software emulator of the Tenstorrent device-level kernel and host APIs. Run tt-metal kernel and host code on a standard x86-64 Linux machine — no Tenstorrent hardware required.
\n\nLinks: Repo
\nofficial · added 2026-07-10
\nemulator, tt-metal, no-hardware, testing, kernels, dev-tools
", "summary": "A C++ software emulator of the Tenstorrent device-level kernel and host APIs. Run tt-metal kernel and host code on a standard x86-64 Linux machine — no Tenstorrent hardware required.", "date_published": "2026-07-10T23:59:36Z", "tags": [ "dev-tools", "official", "entry" ] }, { "id": "https://github.com/moritztng/tt-atom", "url": "https://github.com/moritztng/tt-atom", "title": "tt-atom", "content_html": "Meta's UMA interatomic potential running on Tenstorrent Blackhole — energy, forces, and stress for molecules and periodic materials behind an ASE calculator. Its per-edge Wigner rotation runs as a custom tt-metal kernel for a highest-performance uma-s build.
\n\nLinks: Repo
\nBy @moritztng · affiliated · added 2026-07-10
\nmolecular-dynamics, interatomic-potential, mlip, uma, ase, inference, custom-kernel, ai-models
", "summary": "Meta's UMA interatomic potential running on Tenstorrent Blackhole — energy, forces, and stress for molecules and periodic materials behind an ASE calculator. Its per-edge Wigner rotation runs as a custom tt-metal kernel for a highest-performance uma-s build.", "date_published": "2026-07-10T23:59:35Z", "tags": [ "ai-models", "affiliated", "entry" ] }, { "id": "https://github.com/pcmoritz/libtt", "url": "https://github.com/pcmoritz/libtt", "title": "libtt", "content_html": "A Bazel-built PJRT plugin (libtt.so) providing an XLA backend for Tenstorrent devices. Bundles the tt-xla PJRT implementation with tt-mlir and tt-metal into a single shared object so JAX code runs on Tenstorrent hardware, with patches so sglang-jax works out of the box.
\n\nLinks: Repo
\nBy @Philipp Moritz · community · added 2026-07-10
\nxla, pjrt, jax, bazel, sglang, compilers
", "summary": "A Bazel-built PJRT plugin (libtt.so) providing an XLA backend for Tenstorrent devices. Bundles the tt-xla PJRT implementation with tt-mlir and tt-metal into a single shared object so JAX code runs on Tenstorrent hardware, with patches so sglang-jax works out of the box.", "date_published": "2026-07-10T23:59:34Z", "tags": [ "compilers", "community", "entry" ] }, { "id": "https://github.com/marty1885/ttPseudoRowMajor", "url": "https://github.com/marty1885/ttPseudoRowMajor", "title": "ttPseudoRowMajor", "content_html": "A small TTNN-facing C++ library (ttprm) for running view-shaped tensor work without first materializing the view in DRAM. Targets Tenstorrent TILE tensors and uses cached device operations to gather/scatter through layout views.
\n\nLinks: Repo
\nBy @Martin Chang · community · added 2026-07-10
\nttnn, tensor, tile, layout, cpp, kernels
", "summary": "A small TTNN-facing C++ library (ttprm) for running view-shaped tensor work without first materializing the view in DRAM. Targets Tenstorrent TILE tensors and uses cached device operations to gather/scatter through layout views.", "date_published": "2026-07-10T23:59:33Z", "tags": [ "kernels", "community", "entry" ] }, { "id": "https://github.com/tenstorrent/tt-local-generator/releases/tag/v0.11.0", "url": "https://github.com/tenstorrent/tt-local-generator/releases/tag/v0.11.0", "title": "tt-local-generator v0.11.0", "content_html": "This release brings concurrent job execution to the local art generator, letting you spin up multiple generation tasks without waiting for each to finish, plus adds support for P300X2 image models and improved server discovery for the tt-inference-server integration. The upgrade streamlines workflows for developers working with generative models on Tenstorrent hardware, making it easier to iterate and experiment with batch processing.
\n\nLinks: Repository · v0.11.0
\nofficial · added 2026-07-10
", "summary": "This release brings concurrent job execution to the local art generator, letting you spin up multiple generation tasks without waiting for each to finish, plus adds support for P300X2 image models and improved server discovery for the tt-inference-server integration. The upgrade streamlines workflows for developers working with generative models on Tenstorrent hardware, making it easier to iterate and experiment with batch processing.", "date_published": "2026-07-10T16:47:57Z", "tags": [ "official", "release" ] }, { "id": "https://github.com/tenstorrent/tt-umd/releases/tag/v0.9.9", "url": "https://github.com/tenstorrent/tt-umd/releases/tag/v0.9.9", "title": "tt-umd v0.9.9", "content_html": "This release ships improved device initialization and simulator support. Key fixes include resolving false-positive NOC hang detection that was causing timeouts on healthy devices with tight MMIO budgets, and handling harvested cores correctly so tests don't abort when reading from disabled memory banks on SKUs like Blackhole p100. On the simulator side, the changes unify TLB window setup between RTL and TTSim backends and allow remote devices to accept SocDescriptors directly, laying groundwork for better simulation parity. The release also introduces sccache integration for faster CI builds and adds microbenchmark tests for TTSim clusters.
\n\nLinks: Repository · v0.9.9
\nofficial · added 2026-07-09
", "summary": "This release ships improved device initialization and simulator support. Key fixes include resolving false-positive NOC hang detection that was causing timeouts on healthy devices with tight MMIO budgets, and handling harvested cores correctly so tests don't abort when reading from disabled memory banks on SKUs like Blackhole p100. On the simulator side, the changes unify TLB window setup between RTL and TTSim backends and allow remote devices to accept SocDescriptors directly, laying groundwork for better simulation parity. The release also introduces sccache integration for faster CI builds and adds microbenchmark tests for TTSim clusters.", "date_published": "2026-07-09T17:03:46Z", "tags": [ "official", "release" ] }, { "id": "https://github.com/zk4x/zyx", "url": "https://github.com/zk4x/zyx", "title": "zyx", "content_html": "A complete ML library and compiler in Rust — "from assembly to neural networks" — with a native Tenstorrent backend (src/backend/tenstorrent), autograd, custom kernels, multi-backend support, and Python bindings.
\n\n\nBy @zk4x · community · added 2026-07-03
\nrust, ml-compiler, tensor-library, autograd, backend, compilers, dev-tools
", "summary": "A complete ML library and compiler in Rust — \"from assembly to neural networks\" — with a native Tenstorrent backend (src/backend/tenstorrent), autograd, custom kernels, multi-backend support, and Python bindings.", "date_published": "2026-07-03T23:59:32Z", "tags": [ "compilers", "dev-tools", "community", "entry" ] }, { "id": "https://github.com/RQM-Technologies-dev/tt-rqm-kernels", "url": "https://github.com/RQM-Technologies-dev/tt-rqm-kernels", "title": "tt-rqm-kernels", "content_html": "Structured quaternion, rotor, and phase-aware tensor kernels on ordinary floating-point tensors, plus StructuredBench. Includes CPU/PyTorch references, simulator and emulator paths, and reproducible Wormhole/N300 evidence for quaternion multiply (qmul), fused SU(2) composition, and H2A Hamiltonian lowering.
Links: Repo · Tenstorrent landing page · Wormhole qmul benchmark · Wormhole qmul hardware evidence · SU2ComposeBench · H2A silicon conformance · Current-main qmul port · StructuredBench specification · Tenstorrent RFC
\nBy @RQM-Technologies-dev · community · added 2026-07-03
\nstructured-tensors, quaternion, rotor, tt-metalium, tt-lang, simulator, wormhole, n300, benchmarks, pytorch, custom-kernels, kernels, research
", "summary": "Structured quaternion, rotor, and phase-aware tensor kernels on ordinary floating-point tensors, plus StructuredBench. Includes CPU/PyTorch references, simulator and emulator paths, and reproducible Wormhole/N300 evidence for quaternion multiply (`qmul`), fused SU(2) composition, and H2A Hamiltonian lowering.", "date_published": "2026-07-03T23:59:31Z", "tags": [ "kernels", "research", "community", "entry" ] }, { "id": "https://tenstorrent.github.io/tt-awesome/#cloud-native-support", "url": "https://docs.tenstorrent.com/cloud-native-support/", "title": "Cloud-Native Support", "content_html": "Official documentation hub for running Tenstorrent accelerators on Kubernetes. Centers on tt-operator (the umbrella Helm chart) and covers Node Feature Discovery, kernel-mode driver (tt-kmd) management, firmware flashing, Prometheus telemetry, Fabric Manager topology resolution, Dynamic Resource Allocation, and multi-node scheduling via JobSet and PMIx.
\n\nLinks: docs.tenstorrent.com
\nofficial · added 2026-07-01
\nkubernetes, cloud-native, helm, tt-operator, orchestration, documentation, cloud-infra, guides
", "summary": "Official documentation hub for running Tenstorrent accelerators on Kubernetes. Centers on tt-operator (the umbrella Helm chart) and covers Node Feature Discovery, kernel-mode driver (tt-kmd) management, firmware flashing, Prometheus telemetry, Fabric Manager topology resolution, Dynamic Resource Allocation, and multi-node scheduling via JobSet and PMIx.", "date_published": "2026-07-01T23:59:30Z", "tags": [ "cloud-infra", "guides", "official", "entry" ] }, { "id": "https://github.com/tenstorrent/tt-quietbox2-guide", "url": "https://github.com/tenstorrent/tt-quietbox2-guide", "title": "TT-QuietBox 2 Guide", "content_html": "Official setup and onboarding guide for the TT-QuietBox 2 — a compact, liquid-cooled AI workstation with four Blackhole accelerators, an AMD Ryzen CPU, 256GB RAM, and 4TB NVMe. Covers hardware specs, first-boot setup, and hands-on learning paths for running pre-loaded models like Qwen3-32B and serving text, image, video, and speech models via tt-inference-server.
\n\nLinks: Repo · TT-QuietBox 2 Guide
\nofficial · added 2026-06-30
\nquietbox, blackhole, workstation, setup, getting-started, documentation, hw-system, guides, getting-started
", "summary": "Official setup and onboarding guide for the TT-QuietBox 2 — a compact, liquid-cooled AI workstation with four Blackhole accelerators, an AMD Ryzen CPU, 256GB RAM, and 4TB NVMe. Covers hardware specs, first-boot setup, and hands-on learning paths for running pre-loaded models like Qwen3-32B and serving text, image, video, and speech models via tt-inference-server.", "date_published": "2026-06-30T23:59:29Z", "tags": [ "hw-system", "guides", "getting-started", "official", "entry" ] }, { "id": "https://github.com/tetsuh/tt-metal-community-distro-matrix", "url": "https://github.com/tetsuh/tt-metal-community-distro-matrix", "title": "tetsuh/tt-metal-community-distro-matrix", "content_html": "A compatibility guardrail that continuously monitors whether tt-metal and the official tt-installer build successfully on community Linux distributions that are not part of Tenstorrent's official CI.
\n\nLinks: Repo
\nBy @tetsuh · community · added 2026-06-29
\nhw-system
", "summary": "A compatibility guardrail that continuously monitors whether [tt-metal](https://github.com/tenstorrent/tt-metal) and the official [tt-installer](https://github.com/tenstorrent/tt-installer) build successfully on community Linux distributions that are not part of Tenstorrent's official CI.", "date_published": "2026-06-29T23:59:28Z", "tags": [ "hw-system", "community", "entry" ] }, { "id": "https://github.com/kinginu/tt-splat", "url": "https://github.com/kinginu/tt-splat", "title": "tt-splat — matrix-native 3D Gaussian Splatting on Blackhole", "content_html": "3D Gaussian Splatting rewritten to run on the matrix engine: a polynomial splat and order-independent weighted-sum blending replace exp and depth-sorted alpha, so the pipeline becomes GEMM → activation → GEMM. Renderer + trainer, trained device-resident on a Blackhole p150a.
\n\nLinks: Repo
\nBy @kinginu · community · added 2026-06-29
\n3d-gaussian-splatting, 3dgs, rendering, matrix-engine, weighted-sum-rendering, kernels, research
", "summary": "3D Gaussian Splatting rewritten to run on the matrix engine: a polynomial splat and order-independent weighted-sum blending replace exp and depth-sorted alpha, so the pipeline becomes GEMM → activation → GEMM. Renderer + trainer, trained device-resident on a Blackhole p150a.", "date_published": "2026-06-29T23:59:27Z", "tags": [ "kernels", "research", "community", "entry" ] }, { "id": "https://github.com/tenstorrent/tt-flash/releases/tag/v3.10.0", "url": "https://github.com/tenstorrent/tt-flash/releases/tag/v3.10.0", "title": "tt-flash v3.10.0", "content_html": "This release adds support for Orion SLT silicon, expanding tt-flash's compatibility across Tenstorrent's hardware lineup. If you're working with Orion SLT devices, you can now use tt-flash for firmware management and deployment workflows without workarounds.
\n\nLinks: Repository · v3.10.0
\nofficial · added 2026-06-23
", "summary": "This release adds support for Orion SLT silicon, expanding tt-flash's compatibility across Tenstorrent's hardware lineup. If you're working with Orion SLT devices, you can now use tt-flash for firmware management and deployment workflows without workarounds.", "date_published": "2026-06-23T19:59:38Z", "tags": [ "official", "release" ] }, { "id": "https://github.com/tenstorrent/tt-animatediff/releases/tag/v0.9.0", "url": "https://github.com/tenstorrent/tt-animatediff/releases/tag/v0.9.0", "title": "tt-animatediff v0.9.0", "content_html": "AnimateDiff now includes a Gradio-based web interface alongside full documentation and test coverage, making it more accessible for developers exploring text-to-video generation on Tenstorrent hardware. The addition of Lightning support streamlines integration with existing training workflows, reducing friction for those looking to fine-tune or experiment with animation models at scale.
\n\nLinks: Repository · v0.9.0
\nofficial · added 2026-06-22
", "summary": "AnimateDiff now includes a Gradio-based web interface alongside full documentation and test coverage, making it more accessible for developers exploring text-to-video generation on Tenstorrent hardware. The addition of Lightning support streamlines integration with existing training workflows, reducing friction for those looking to fine-tune or experiment with animation models at scale.", "date_published": "2026-06-22T19:56:31Z", "tags": [ "official", "release" ] }, { "id": "https://github.com/tenstorrent/ttsim-qemu", "url": "https://github.com/tenstorrent/ttsim-qemu", "title": "ttsim-qemu", "content_html": "Tenstorrent's fork of QEMU that provides the full-system emulation layer behind ttsim. Models the RISC-V cores and system devices of Wormhole and Blackhole so TT-Metalium workloads can boot and run without physical silicon.
\n\nLinks: Repo
\nofficial · added 2026-06-16
\nsimulator, qemu, full-system, emulation, no-hardware, wormhole, blackhole, riscv-arch, dev-tools
", "summary": "Tenstorrent's fork of QEMU that provides the full-system emulation layer behind ttsim. Models the RISC-V cores and system devices of Wormhole and Blackhole so TT-Metalium workloads can boot and run without physical silicon.", "date_published": "2026-06-16T23:59:26Z", "tags": [ "riscv-arch", "dev-tools", "official", "entry" ] }, { "id": "https://github.com/tenstorrent/tt-forge-onnx", "url": "https://github.com/tenstorrent/tt-forge-onnx", "title": "tt-forge-onnx", "content_html": "ONNX graph compiler for Tenstorrent hardware. Optimizes and transforms ONNX model graphs for efficient execution on Tensix accelerators. Used as a backend by tt-forge for ONNX model ingestion.
\n\nLinks: Repo
\nofficial · added 2026-06-15
\nonnx, compiler, graph-optimization, mlir, compilers
", "summary": "ONNX graph compiler for Tenstorrent hardware. Optimizes and transforms ONNX model graphs for efficient execution on Tensix accelerators. Used as a backend by tt-forge for ONNX model ingestion.", "date_published": "2026-06-15T23:59:25Z", "tags": [ "compilers", "official", "entry" ] }, { "id": "https://github.com/tenstorrent/tt-animatediff", "url": "https://github.com/tenstorrent/tt-animatediff", "title": "tt-animatediff", "content_html": "Generates short, temporally coherent animated GIFs using the AnimateDiff model on Tenstorrent hardware. Phase 1 runs the correct SD 1.4 + MotionAdapter architecture on CPU; Phase 2 accelerates spatial denoising on Blackhole using the TTNN UNet. Produces vibrant 8-frame animations in ~15 s/frame on a P300C.
\n\nLinks: Repo · Native Video Animation with AnimateDiff (VSCode Toolkit)
\nofficial · added 2026-06-03
\nanimatediff, video-generation, stable-diffusion, diffusion, gif, blackhole, ai-models, games-demos
", "summary": "Generates short, temporally coherent animated GIFs using the AnimateDiff model on Tenstorrent hardware. Phase 1 runs the correct SD 1.4 + MotionAdapter architecture on CPU; Phase 2 accelerates spatial denoising on Blackhole using the TTNN UNet. Produces vibrant 8-frame animations in ~15 s/frame on a P300C.", "date_published": "2026-06-03T23:59:24Z", "tags": [ "ai-models", "games-demos", "official", "entry" ] }, { "id": "https://docs.tenstorrent.com/tt-vscode-toolkit/lessons/animatediff-video-generation/", "url": "https://docs.tenstorrent.com/tt-vscode-toolkit/lessons/animatediff-video-generation/", "title": "tt-animatediff — Native Video Animation with AnimateDiff (VSCode Toolkit)", "content_html": "Generates short, temporally coherent animated GIFs using the AnimateDiff model on Tenstorrent hardware. Phase 1 runs the correct SD 1.4 + MotionAdapter architecture on CPU; Phase 2 accelerates spatial denoising on Blackhole using the TTNN UNet. Produces vibrant 8-frame animations in ~15 s/frame on a P300C.
\n\nLinks: Native Video Animation with AnimateDiff (VSCode Toolkit)
\nofficial · added 2026-06-03
\nanimatediff, video-generation, stable-diffusion, diffusion, gif, blackhole, ai-models, games-demos
", "summary": "Generates short, temporally coherent animated GIFs using the AnimateDiff model on Tenstorrent hardware. Phase 1 runs the correct SD 1.4 + MotionAdapter architecture on CPU; Phase 2 accelerates spatial denoising on Blackhole using the TTNN UNet. Produces vibrant 8-frame animations in ~15 s/frame on a P300C.", "date_published": "2026-06-03T23:59:24Z", "tags": [ "ai-models", "games-demos", "official", "lesson", "article" ] }, { "id": "https://github.com/Zaneham/ttas/releases/tag/v0.1.0", "url": "https://github.com/Zaneham/ttas/releases/tag/v0.1.0", "title": "ttas v0.1.0", "content_html": "ttas v0.1.0 reaches a significant milestone: all 128 Wormhole b0 instructions are now fully cross-checked against Tenstorrent's canonical TT_OP_ macros in ckernel_ops.h, verifying opcodes, start bits, and field widths across the complete ISA. There's one breaking change — positional operand order in .tts files now follows the TT_OP_ macro signature rather than ascending start_bit order; named operands are unaffected.
\n\nLinks: Repository · v0.1.0
\ncommunity · added 2026-05-28
", "summary": "ttas v0.1.0 reaches a significant milestone: all 128 Wormhole b0 instructions are now fully cross-checked against Tenstorrent's canonical TT_OP_ macros in ckernel_ops.h, verifying opcodes, start bits, and field widths across the complete ISA. There's one breaking change — positional operand order in .tts files now follows the TT_OP_ macro signature rather than ascending start_bit order; named operands are unaffected.", "date_published": "2026-05-28T07:08:35Z", "tags": [ "community", "release" ] }, { "id": "https://github.com/Zaneham/ttas", "url": "https://github.com/Zaneham/ttas", "title": "ttas", "content_html": "ttas is a hacker-friendly assembler/disassembler for Tensix on Wormhole. It turns assembly into the exact 32-bit words the hardware runs, and turns binaries back into readable instructions using the same shared instruction table.
\n\nLinks: Repo
\nBy @Zaneham · community · added 2026-05-27
\nassembler, dev-tools, hw-system
", "summary": "ttas is a hacker-friendly assembler/disassembler for Tensix on Wormhole. It turns assembly into the exact 32-bit words the hardware runs, and turns binaries back into readable instructions using the same shared instruction table.", "date_published": "2026-05-27T23:59:23Z", "tags": [ "dev-tools", "hw-system", "community", "entry" ] }, { "id": "https://github.com/Knight-Ops/libtt-metal-cxx", "url": "https://github.com/Knight-Ops/libtt-metal-cxx", "title": "libtt-metal-cxx", "content_html": "Rust crate that exposes the TT-Metal host API through a C++ bridge via cxx.rs — covering device management, program/kernel creation (from source file or inline string), circular buffers, semaphores, runtime arguments, sharded buffers, and MeshDevice workflows, with hardware-backed integration tests.
\n\nLinks: Repo
\nBy @Knight-Ops · community · added 2026-05-20
\nrust, bindings, cxx, tt-metal, ffi, host-api, dev-tools
", "summary": "Rust crate that exposes the TT-Metal host API through a C++ bridge via cxx.rs — covering device management, program/kernel creation (from source file or inline string), circular buffers, semaphores, runtime arguments, sharded buffers, and MeshDevice workflows, with hardware-backed integration tests.", "date_published": "2026-05-20T23:59:22Z", "tags": [ "dev-tools", "community", "entry" ] }, { "id": "https://github.com/tenstorrent/tt-llk", "url": "https://github.com/tenstorrent/tt-llk", "title": "tt-llk", "content_html": "Tenstorrent Low-Level Kernels: the C++ library that directly programs the RISC-V cores inside each Tensix compute engine. TRISC0 (unpack), TRISC1 (math/FPU/SFPU), and TRISC2 (pack) are all programmed through this layer — it is the interface between TT-Metal kernel code and bare silicon.
\n\nLinks: Repo · Top-level architecture overview
\nofficial · added 2026-05-13
\ntensix, risc-v, llk, trisc, brisc, ncrisc, low-level, compute-engine, kernels, riscv-arch
", "summary": "Tenstorrent Low-Level Kernels: the C++ library that directly programs the RISC-V cores inside each Tensix compute engine. TRISC0 (unpack), TRISC1 (math/FPU/SFPU), and TRISC2 (pack) are all programmed through this layer — it is the interface between TT-Metal kernel code and bare silicon.", "date_published": "2026-05-13T23:59:21Z", "tags": [ "kernels", "riscv-arch", "official", "entry" ] }, { "id": "https://github.com/tenstorrent/tt-llk/blob/main/docs/llk/l2/top_level_overview.md", "url": "https://github.com/tenstorrent/tt-llk/blob/main/docs/llk/l2/top_level_overview.md", "title": "tt-llk — Top-level architecture overview", "content_html": "Tenstorrent Low-Level Kernels: the C++ library that directly programs the RISC-V cores inside each Tensix compute engine. TRISC0 (unpack), TRISC1 (math/FPU/SFPU), and TRISC2 (pack) are all programmed through this layer — it is the interface between TT-Metal kernel code and bare silicon.
\n\nLinks: Top-level architecture overview
\nofficial · added 2026-05-13
\ntensix, risc-v, llk, trisc, brisc, ncrisc, low-level, compute-engine, kernels, riscv-arch
", "summary": "Tenstorrent Low-Level Kernels: the C++ library that directly programs the RISC-V cores inside each Tensix compute engine. TRISC0 (unpack), TRISC1 (math/FPU/SFPU), and TRISC2 (pack) are all programmed through this layer — it is the interface between TT-Metal kernel code and bare silicon.", "date_published": "2026-05-13T23:59:21Z", "tags": [ "kernels", "riscv-arch", "official", "article", "article" ] }, { "id": "https://tenstorrent.github.io/tt-awesome/#lecture-wm-csci654-tenstorrent", "url": "https://www.youtube.com/watch?v=CixEFPc8oxg", "title": "Tenstorrent Architecture — W&M CSCI654 Advanced Computer Architecture", "content_html": "Lecture 20 from William & Mary's graduate Computer Architecture course. Frames Tenstorrent in the landscape between GPUs and TPUs, draws comparisons to Cerebras and SambaNova, then dives deep into the Wormhole chip and Tensix core: the 5 RISC-V core design, SFPU, NoC, and dataflow execution model.
\n\nLinks: Lecture 20 — Tenstorrent Architecture (YouTube)
\nBy Yifan & GPU / William & Mary · community · added 2026-05-13
\nlecture, architecture, wormhole, tensix, risc-v, sfpu, noc, academia, guides, riscv-arch
", "summary": "Lecture 20 from William & Mary's graduate Computer Architecture course. Frames Tenstorrent in the landscape between GPUs and TPUs, draws comparisons to Cerebras and SambaNova, then dives deep into the Wormhole chip and Tensix core: the 5 RISC-V core design, SFPU, NoC, and dataflow execution model.", "date_published": "2026-05-13T23:59:20Z", "tags": [ "guides", "riscv-arch", "community", "entry" ] }, { "id": "https://www.youtube.com/watch?v=CixEFPc8oxg", "url": "https://www.youtube.com/watch?v=CixEFPc8oxg", "title": "Tenstorrent Architecture — W&M CSCI654 Advanced Computer Architecture — Lecture 20 — Tenstorrent Architecture (YouTube)", "content_html": "Lecture 20 from William & Mary's graduate Computer Architecture course. Frames Tenstorrent in the landscape between GPUs and TPUs, draws comparisons to Cerebras and SambaNova, then dives deep into the Wormhole chip and Tensix core: the 5 RISC-V core design, SFPU, NoC, and dataflow execution model.
\n\nLinks: Lecture 20 — Tenstorrent Architecture (YouTube)
\nBy Yifan & GPU / William & Mary · community · added 2026-05-13
\nlecture, architecture, wormhole, tensix, risc-v, sfpu, noc, academia, guides, riscv-arch
", "summary": "Lecture 20 from William & Mary's graduate Computer Architecture course. Frames Tenstorrent in the landscape between GPUs and TPUs, draws comparisons to Cerebras and SambaNova, then dives deep into the Wormhole chip and Tensix core: the 5 RISC-V core design, SFPU, NoC, and dataflow execution model.", "date_published": "2026-05-13T23:59:20Z", "tags": [ "guides", "riscv-arch", "community", "video", "article" ] }, { "id": "https://tenstorrent.github.io/tt-awesome/#paper-attention-grayskull", "url": "https://arxiv.org/abs/2407.13885", "title": "Attention in SRAM on Tenstorrent Grayskull", "content_html": "A fused kernel for the Grayskull architecture implementing Transformer self-attention entirely within SRAM. Combines matrix multiply, attention score scaling, and Softmax without DRAM accesses, achieving significant speedups over non-fused implementations.
\n\nLinks: arXiv:2407.13885
\nBy Moritz Thüning · community · added 2026-05-13
\nattention, transformer, sram, grayskull, kernel, risc-v, research, kernels
", "summary": "A fused kernel for the Grayskull architecture implementing Transformer self-attention entirely within SRAM. Combines matrix multiply, attention score scaling, and Softmax without DRAM accesses, achieving significant speedups over non-fused implementations.", "date_published": "2026-05-13T23:59:19Z", "tags": [ "research", "kernels", "community", "entry" ] }, { "id": "https://arxiv.org/abs/2407.13885", "url": "https://arxiv.org/abs/2407.13885", "title": "Attention in SRAM on Tenstorrent Grayskull — arXiv:2407.13885", "content_html": "A fused kernel for the Grayskull architecture implementing Transformer self-attention entirely within SRAM. Combines matrix multiply, attention score scaling, and Softmax without DRAM accesses, achieving significant speedups over non-fused implementations.
\n\nLinks: arXiv:2407.13885
\nBy Moritz Thüning · community · added 2026-05-13
\nattention, transformer, sram, grayskull, kernel, risc-v, research, kernels
", "summary": "A fused kernel for the Grayskull architecture implementing Transformer self-attention entirely within SRAM. Combines matrix multiply, attention score scaling, and Softmax without DRAM accesses, achieving significant speedups over non-fused implementations.", "date_published": "2026-05-13T23:59:19Z", "tags": [ "research", "kernels", "community", "paper", "article" ] }, { "id": "https://tenstorrent.github.io/tt-awesome/#paper-matmul-grayskull", "url": "https://arxiv.org/abs/2505.06085", "title": "Assessing Tenstorrent Grayskull RISC-V MatMul Acceleration for LLMs", "content_html": "Evaluates the Tenstorrent Grayskull e75 RISC-V accelerator for matrix multiplication at reduced numerical precision (BFP8 and LoFi), a fundamental kernel in LLM inference computation.
\n\nLinks: arXiv:2505.06085
\nBy Hiari Pizzini Cavagna, Daniele Cesarini, Andrea Bartolini · community · added 2026-05-13
\nmatmul, grayskull, risc-v, bfp8, lofi, llm, precision, research
", "summary": "Evaluates the Tenstorrent Grayskull e75 RISC-V accelerator for matrix multiplication at reduced numerical precision (BFP8 and LoFi), a fundamental kernel in LLM inference computation.", "date_published": "2026-05-13T23:59:18Z", "tags": [ "research", "community", "entry" ] }, { "id": "https://arxiv.org/abs/2505.06085", "url": "https://arxiv.org/abs/2505.06085", "title": "Assessing Tenstorrent Grayskull RISC-V MatMul Acceleration for LLMs — arXiv:2505.06085", "content_html": "Evaluates the Tenstorrent Grayskull e75 RISC-V accelerator for matrix multiplication at reduced numerical precision (BFP8 and LoFi), a fundamental kernel in LLM inference computation.
\n\nLinks: arXiv:2505.06085
\nBy Hiari Pizzini Cavagna, Daniele Cesarini, Andrea Bartolini · community · added 2026-05-13
\nmatmul, grayskull, risc-v, bfp8, lofi, llm, precision, research
", "summary": "Evaluates the Tenstorrent Grayskull e75 RISC-V accelerator for matrix multiplication at reduced numerical precision (BFP8 and LoFi), a fundamental kernel in LLM inference computation.", "date_published": "2026-05-13T23:59:18Z", "tags": [ "research", "community", "paper", "article" ] }, { "id": "https://tenstorrent.github.io/tt-awesome/#paper-nbody-strategies-wormhole", "url": "https://arxiv.org/abs/2605.02744", "title": "Porting Strategies for Gravitational N-Body Simulations on Tenstorrent Wormhole", "content_html": "Evaluates three strategies for scaling an N-body code across multiple Tenstorrent Wormhole accelerators. Builds on the established performance of single-card N-body work to explore parallelism via the on-chip NoC and multi-accelerator configurations.
\n\nLinks: arXiv:2605.02744
\nBy Jenny Lynn Almerol, Elisabetta Boella, Mario Spera, Daniele Gregori · community · added 2026-05-13
\nn-body, astrophysics, hpc, wormhole, risc-v, multi-accelerator, simulation, research
", "summary": "Evaluates three strategies for scaling an N-body code across multiple Tenstorrent Wormhole accelerators. Builds on the established performance of single-card N-body work to explore parallelism via the on-chip NoC and multi-accelerator configurations.", "date_published": "2026-05-13T23:59:17Z", "tags": [ "research", "community", "entry" ] }, { "id": "https://arxiv.org/abs/2605.02744", "url": "https://arxiv.org/abs/2605.02744", "title": "Porting Strategies for Gravitational N-Body Simulations on Tenstorrent Wormhole — arXiv:2605.02744", "content_html": "Evaluates three strategies for scaling an N-body code across multiple Tenstorrent Wormhole accelerators. Builds on the established performance of single-card N-body work to explore parallelism via the on-chip NoC and multi-accelerator configurations.
\n\nLinks: arXiv:2605.02744
\nBy Jenny Lynn Almerol, Elisabetta Boella, Mario Spera, Daniele Gregori · community · added 2026-05-13
\nn-body, astrophysics, hpc, wormhole, risc-v, multi-accelerator, simulation, research
", "summary": "Evaluates three strategies for scaling an N-body code across multiple Tenstorrent Wormhole accelerators. Builds on the established performance of single-card N-body work to explore parallelism via the on-chip NoC and multi-accelerator configurations.", "date_published": "2026-05-13T23:59:17Z", "tags": [ "research", "community", "paper", "article" ] }, { "id": "https://tenstorrent.github.io/tt-awesome/#paper-tileloom", "url": "https://arxiv.org/abs/2512.22168", "title": "TileLoom: Automatic Dataflow Planning for Spatial Dataflow Accelerators", "content_html": "Compiler system that automatically generates efficient dataflow plans for tile-based languages on spatial accelerators including Tenstorrent Wormhole. Exploits on-chip network forwarding between processing elements to reduce DRAM pressure.
\n\nLinks: arXiv:2512.22168
\nBy Wei Li, Zhenyu Bai, Heru Wang, Pranav Dangi · community · added 2026-05-13
\ncompiler, dataflow, spatial-accelerator, tile-based, on-chip-network, wormhole, research, compilers
", "summary": "Compiler system that automatically generates efficient dataflow plans for tile-based languages on spatial accelerators including Tenstorrent Wormhole. Exploits on-chip network forwarding between processing elements to reduce DRAM pressure.", "date_published": "2026-05-13T23:59:16Z", "tags": [ "research", "compilers", "community", "entry" ] }, { "id": "https://arxiv.org/abs/2512.22168", "url": "https://arxiv.org/abs/2512.22168", "title": "TileLoom: Automatic Dataflow Planning for Spatial Dataflow Accelerators — arXiv:2512.22168", "content_html": "Compiler system that automatically generates efficient dataflow plans for tile-based languages on spatial accelerators including Tenstorrent Wormhole. Exploits on-chip network forwarding between processing elements to reduce DRAM pressure.
\n\nLinks: arXiv:2512.22168
\nBy Wei Li, Zhenyu Bai, Heru Wang, Pranav Dangi · community · added 2026-05-13
\ncompiler, dataflow, spatial-accelerator, tile-based, on-chip-network, wormhole, research, compilers
", "summary": "Compiler system that automatically generates efficient dataflow plans for tile-based languages on spatial accelerators including Tenstorrent Wormhole. Exploits on-chip network forwarding between processing elements to reduce DRAM pressure.", "date_published": "2026-05-13T23:59:16Z", "tags": [ "research", "compilers", "community", "paper", "article" ] }, { "id": "https://tenstorrent.github.io/tt-awesome/#paper-tts-lightning", "url": "https://arxiv.org/abs/2604.03279", "title": "Rewriting TTS Inference Economics: Lightning V2 on Tenstorrent vs. NVIDIA L40S", "content_html": "Shows that Text-to-Speech inference on Tenstorrent Lightning V2 achieves 4× lower cost than NVIDIA L40S. Applies BlockFloat8 (BFP8) and low-fidelity (LoFi) precision strategies to TTS despite their greater numerical fragility compared to LLMs.
\n\nLinks: arXiv:2604.03279
\nBy Ranjith M. S., Akshat Mandloi, Sudarshan Kamath · community · added 2026-05-13
\ntts, text-to-speech, inference, bfp8, lofi, cost-efficiency, precision, research, ai-models
", "summary": "Shows that Text-to-Speech inference on Tenstorrent Lightning V2 achieves 4× lower cost than NVIDIA L40S. Applies BlockFloat8 (BFP8) and low-fidelity (LoFi) precision strategies to TTS despite their greater numerical fragility compared to LLMs.", "date_published": "2026-05-13T23:59:15Z", "tags": [ "research", "ai-models", "community", "entry" ] }, { "id": "https://arxiv.org/abs/2604.03279", "url": "https://arxiv.org/abs/2604.03279", "title": "Rewriting TTS Inference Economics: Lightning V2 on Tenstorrent vs. NVIDIA L40S — arXiv:2604.03279", "content_html": "Shows that Text-to-Speech inference on Tenstorrent Lightning V2 achieves 4× lower cost than NVIDIA L40S. Applies BlockFloat8 (BFP8) and low-fidelity (LoFi) precision strategies to TTS despite their greater numerical fragility compared to LLMs.
\n\nLinks: arXiv:2604.03279
\nBy Ranjith M. S., Akshat Mandloi, Sudarshan Kamath · community · added 2026-05-13
\ntts, text-to-speech, inference, bfp8, lofi, cost-efficiency, precision, research, ai-models
", "summary": "Shows that Text-to-Speech inference on Tenstorrent Lightning V2 achieves 4× lower cost than NVIDIA L40S. Applies BlockFloat8 (BFP8) and low-fidelity (LoFi) precision strategies to TTS despite their greater numerical fragility compared to LLMs.", "date_published": "2026-05-13T23:59:15Z", "tags": [ "research", "ai-models", "community", "paper", "article" ] }, { "id": "https://tenstorrent.github.io/tt-awesome/#blog-anuraagw-blackhole-arch", "url": "https://anuraagw.me/blog/blackhole-architecture", "title": "Tenstorrent Blackhole Architecture Guide", "content_html": "A 6,500-word community deep dive into the Blackhole p100a architecture: the tile model (Tensix, DRAM, SiFive x280 L2CPU, Ethernet, PCIe, NoC arc), firmware startup sequence, MOP micro-op processor, replay buffer, FPU/SFPU sync, and the anatomy of a kernel. From the author of blackhole-py.
\n\nLinks: anuraagw.me — February 2026
\nBy boopdotpng · community · added 2026-05-13
\nblackhole, architecture, tensix, noc, sifive-x280, firmware, mop, sfpu, deep-dive, blog, riscv-arch, guides, blogs
", "summary": "A 6,500-word community deep dive into the Blackhole p100a architecture: the tile model (Tensix, DRAM, SiFive x280 L2CPU, Ethernet, PCIe, NoC arc), firmware startup sequence, MOP micro-op processor, replay buffer, FPU/SFPU sync, and the anatomy of a kernel. From the author of blackhole-py.", "date_published": "2026-05-13T23:59:14Z", "tags": [ "riscv-arch", "guides", "blogs", "community", "entry" ] }, { "id": "https://anuraagw.me/blog/blackhole-architecture", "url": "https://anuraagw.me/blog/blackhole-architecture", "title": "Tenstorrent Blackhole Architecture Guide — anuraagw.me — February 2026", "content_html": "A 6,500-word community deep dive into the Blackhole p100a architecture: the tile model (Tensix, DRAM, SiFive x280 L2CPU, Ethernet, PCIe, NoC arc), firmware startup sequence, MOP micro-op processor, replay buffer, FPU/SFPU sync, and the anatomy of a kernel. From the author of blackhole-py.
\n\nLinks: anuraagw.me — February 2026
\nBy boopdotpng · community · added 2026-05-13
\nblackhole, architecture, tensix, noc, sifive-x280, firmware, mop, sfpu, deep-dive, blog, riscv-arch, guides, blogs
", "summary": "A 6,500-word community deep dive into the Blackhole p100a architecture: the tile model (Tensix, DRAM, SiFive x280 L2CPU, Ethernet, PCIe, NoC arc), firmware startup sequence, MOP micro-op processor, replay buffer, FPU/SFPU sync, and the anatomy of a kernel. From the author of blackhole-py.", "date_published": "2026-05-13T23:59:14Z", "tags": [ "riscv-arch", "guides", "blogs", "community", "article", "article" ] }, { "id": "https://tenstorrent.github.io/tt-awesome/#paper-stencils-wormhole", "url": "https://arxiv.org/abs/2605.07599", "title": "Stencil Computations on Tenstorrent Wormhole", "content_html": "Maps 2D 5-point stencil computations onto the Tenstorrent Wormhole RISC-V AI dataflow accelerator via two implementations: element-wise decomposition (Axpy) and matrix-multiplication reformulation (MatMul). Profiling shows the isolated Wormhole kernel is competitive with CPU execution, with PCIe transfers and initialization driving end-to-end overhead; Axpy achieves lower energy than the CPU baseline at large scales. Identifies architectural and software directions for making AI accelerators viable for HPC stencil workloads. 2025.
\n\nLinks: arXiv:2605.07599
\ncommunity · added 2026-05-12
\nstencil, hpc, wormhole, risc-v, energy-efficiency, benchmarks, dataflow, research
", "summary": "Maps 2D 5-point stencil computations onto the Tenstorrent Wormhole RISC-V AI dataflow accelerator via two implementations: element-wise decomposition (Axpy) and matrix-multiplication reformulation (MatMul). Profiling shows the isolated Wormhole kernel is competitive with CPU execution, with PCIe transfers and initialization driving end-to-end overhead; Axpy achieves lower energy than the CPU baseline at large scales. Identifies architectural and software directions for making AI accelerators viable for HPC stencil workloads. 2025.", "date_published": "2026-05-12T23:59:13Z", "tags": [ "research", "community", "entry" ] }, { "id": "https://arxiv.org/abs/2605.07599", "url": "https://arxiv.org/abs/2605.07599", "title": "Stencil Computations on Tenstorrent Wormhole — arXiv:2605.07599", "content_html": "Maps 2D 5-point stencil computations onto the Tenstorrent Wormhole RISC-V AI dataflow accelerator via two implementations: element-wise decomposition (Axpy) and matrix-multiplication reformulation (MatMul). Profiling shows the isolated Wormhole kernel is competitive with CPU execution, with PCIe transfers and initialization driving end-to-end overhead; Axpy achieves lower energy than the CPU baseline at large scales. Identifies architectural and software directions for making AI accelerators viable for HPC stencil workloads. 2025.
\n\nLinks: arXiv:2605.07599
\ncommunity · added 2026-05-12
\nstencil, hpc, wormhole, risc-v, energy-efficiency, benchmarks, dataflow, research
", "summary": "Maps 2D 5-point stencil computations onto the Tenstorrent Wormhole RISC-V AI dataflow accelerator via two implementations: element-wise decomposition (Axpy) and matrix-multiplication reformulation (MatMul). Profiling shows the isolated Wormhole kernel is competitive with CPU execution, with PCIe transfers and initialization driving end-to-end overhead; Axpy achieves lower energy than the CPU baseline at large scales. Identifies architectural and software directions for making AI accelerators viable for HPC stencil workloads. 2025.", "date_published": "2026-05-12T23:59:13Z", "tags": [ "research", "community", "paper", "article" ] }, { "id": "https://tenstorrent.github.io/tt-awesome/#tt-console", "url": "https://console.tenstorrent.com", "title": "TT Console", "content_html": "Browser-based cloud console for exploring AI on Tenstorrent hardware. Run LLM inference, image and video generation, and browse the supported model catalog in-browser — backed by Tenstorrent accelerators. Cloud hardware access and advanced workflows (deployments, agents) available in staged rollout.
\n\nLinks: console.tenstorrent.com
\nofficial · added 2026-05-11
\ncloud, console, inference, playground, llm, image-generation, video-generation, demo, cloud-infra, ai-models
", "summary": "Browser-based cloud console for exploring AI on Tenstorrent hardware. Run LLM inference, image and video generation, and browse the supported model catalog in-browser — backed by Tenstorrent accelerators. Cloud hardware access and advanced workflows (deployments, agents) available in staged rollout.", "date_published": "2026-05-11T23:59:12Z", "tags": [ "cloud-infra", "ai-models", "official", "entry" ] }, { "id": "https://github.com/tenstorrent/whisper/releases/tag/1.861", "url": "https://github.com/tenstorrent/whisper/releases/tag/1.861", "title": "whisper 1.861", "content_html": "whisper released 1.861. Repository: https://github.com/tenstorrent/whisper
\n\nLinks: Repository · 1.861
\nofficial · added 2026-05-11
", "summary": "whisper released 1.861. Repository: https://github.com/tenstorrent/whisper", "date_published": "2026-05-11T15:44:36Z", "tags": [ "official", "release" ] }, { "id": "https://github.com/mesham/tt-sim/releases/tag/v1.0", "url": "https://github.com/mesham/tt-sim/releases/tag/v1.0", "title": "tt-sim v1.0", "content_html": "tt-sim v1.0 is the first tagged release of a community-built Tenstorrent hardware simulator, providing a simulation environment compatible with the mesham/tt-metal fork. This gives developers an accessible way to explore tt-metal on Tenstorrent architecture without requiring physical hardware.
\n\nLinks: Repository · v1.0
\ncommunity · added 2026-05-11
", "summary": "tt-sim v1.0 is the first tagged release of a community-built Tenstorrent hardware simulator, providing a simulation environment compatible with the mesham/tt-metal fork. This gives developers an accessible way to explore tt-metal on Tenstorrent architecture without requiring physical hardware.", "date_published": "2026-05-11T13:07:42Z", "tags": [ "community", "release" ] }, { "id": "https://github.com/tenstorrent/tt-metal", "url": "https://github.com/tenstorrent/tt-metal", "title": "tt-metal", "content_html": "TT-NN operator library and TT-Metalium low-level kernel programming model. The primary SDK for developing on Tenstorrent hardware — from high-level tensor ops to bare-metal RISC-V kernels.
\n\n\nofficial · added 2026-05-08
\nmetalium, ttnn, sdk, kernels, core, kernels, compilers, getting-started
", "summary": "TT-NN operator library and TT-Metalium low-level kernel programming model. The primary SDK for developing on Tenstorrent hardware — from high-level tensor ops to bare-metal RISC-V kernels.", "date_published": "2026-05-08T23:59:11Z", "tags": [ "kernels", "compilers", "getting-started", "official", "entry" ] }, { "id": "https://github.com/tenstorrent/tt-forge", "url": "https://github.com/tenstorrent/tt-forge", "title": "tt-forge", "content_html": "Tenstorrent's MLIR-based compiler frontend. Enables running AI workloads from PyTorch, ONNX, and other frameworks on all Tenstorrent hardware configurations through an open-source, general, and performant compiler.
\n\n\nofficial · added 2026-05-08
\nmlir, compiler, pytorch, onnx, frontend, compilers, getting-started
", "summary": "Tenstorrent's MLIR-based compiler frontend. Enables running AI workloads from PyTorch, ONNX, and other frameworks on all Tenstorrent hardware configurations through an open-source, general, and performant compiler.", "date_published": "2026-05-08T23:59:10Z", "tags": [ "compilers", "getting-started", "official", "entry" ] }, { "id": "https://github.com/tenstorrent-riscv-software/tt-bh-linux/releases/tag/v0.11", "url": "https://github.com/tenstorrent-riscv-software/tt-bh-linux/releases/tag/v0.11", "title": "tt-bh-linux v0.11", "content_html": "This release upgrades to Linux Kernel 7.0 and adds device tree support for performance monitoring, letting you track system metrics directly from userspace. The kernel now ships with a patch that eliminates spurious swiotlb errors during boot, cleaning up initialization logs and making debugging easier. Several fixes to the console tool and build system round out the update.
\n\nLinks: Repository · v0.11
\nofficial · added 2026-04-13
", "summary": "This release upgrades to Linux Kernel 7.0 and adds device tree support for performance monitoring, letting you track system metrics directly from userspace. The kernel now ships with a patch that eliminates spurious swiotlb errors during boot, cleaning up initialization logs and making debugging easier. Several fixes to the console tool and build system round out the update.", "date_published": "2026-04-13T15:10:59Z", "tags": [ "official", "release" ] }, { "id": "https://github.com/tenstorrent/tt-firmware/releases/tag/v19.6.0", "url": "https://github.com/tenstorrent/tt-firmware/releases/tag/v19.6.0", "title": "tt-firmware v19.6.0", "content_html": "This release brings memory and stability improvements across Tenstorrent's platform lineup, with Wormhole now supporting Samsung GDDR6 and both platforms implementing security wipes during initialization. On Blackhole, the p300 gains critical fixes for reboot reliability and expanded board support with Galaxy revC, while p300c power limits have been tuned (550W board power, 125W TDP, 88°C GDDR thermals). The telemetry layer also gets new visibility into AICLK arbitration decisions with granular tracing, helping developers debug frequency scaling behavior more effectively.
\n\nLinks: Repository · v19.6.0
\nofficial · added 2026-02-20
", "summary": "This release brings memory and stability improvements across Tenstorrent's platform lineup, with Wormhole now supporting Samsung GDDR6 and both platforms implementing security wipes during initialization. On Blackhole, the p300 gains critical fixes for reboot reliability and expanded board support with Galaxy revC, while p300c power limits have been tuned (550W board power, 125W TDP, 88°C GDDR thermals). The telemetry layer also gets new visibility into AICLK arbitration decisions with granular tracing, helping developers debug frequency scaling behavior more effectively.", "date_published": "2026-02-20T16:53:34Z", "tags": [ "official", "release" ] }, { "id": "https://github.com/Syllo/nvtop/releases/tag/3.3.2", "url": "https://github.com/Syllo/nvtop/releases/tag/3.3.2", "title": "nvtop 3.3.2", "content_html": "This patch release fixes snapshot mode to generate valid JSON output, addressing a parsing issue that would have affected downstream tooling. The update also introduces loop snapshot mode for continuous monitoring and extends snapshot data to capture process-level details and encode/decode statistics, giving users more granular visibility into system activity.
\n\nLinks: Repository · 3.3.2
\ncommunity · added 2026-02-08
", "summary": "This patch release fixes snapshot mode to generate valid JSON output, addressing a parsing issue that would have affected downstream tooling. The update also introduces loop snapshot mode for continuous monitoring and extends snapshot data to capture process-level details and encode/decode statistics, giving users more granular visibility into system activity.", "date_published": "2026-02-08T17:57:16Z", "tags": [ "community", "release" ] }, { "id": "https://github.com/tenstorrent/tt-tools-common/releases/tag/v1.6.0", "url": "https://github.com/tenstorrent/tt-tools-common/releases/tag/v1.6.0", "title": "tt-tools-common v1.6.0", "content_html": "A packaging-focused cut: beyond the version bump, this build generates proper Ubuntu and Debian changelogs — it's the version now shipping as python3-tt-tools-common in the Tenstorrent PPA, so the shared plumbing under tt-smi, tt-flash, and friends installs cleanly via apt.
Links: Repository · v1.6.0
\nofficial · added 2025-12-23
", "summary": "A packaging-focused cut: beyond the version bump, this build generates proper Ubuntu and Debian changelogs — it's the version now shipping as `python3-tt-tools-common` in the [Tenstorrent PPA](https://ppa.tenstorrent.com/), so the shared plumbing under tt-smi, tt-flash, and friends installs cleanly via apt.", "date_published": "2025-12-23T21:02:08Z", "tags": [ "official", "release" ] }, { "id": "https://github.com/tenstorrent/tt-system-tools/releases/tag/v1.4.1", "url": "https://github.com/tenstorrent/tt-system-tools/releases/tag/v1.4.1", "title": "tt-system-tools v1.4.1", "content_html": "tt-system-tools released v1.4.1. Repository: https://github.com/tenstorrent/tt-system-tools
\n\nLinks: Repository · v1.4.1
\nofficial · added 2025-12-08
", "summary": "tt-system-tools released v1.4.1. Repository: https://github.com/tenstorrent/tt-system-tools", "date_published": "2025-12-08T17:23:48Z", "tags": [ "official", "release" ] }, { "id": "https://github.com/tenstorrent/riescue/releases/tag/v1.7.0", "url": "https://github.com/tenstorrent/riescue/releases/tag/v1.7.0", "title": "RiESCUE v1.7.0", "content_html": "The loader and scheduler in RiescueD have been refactored with explicit interfaces that make their control flow clearer and reduce redundant checks—test_setup now runs once during scheduler initialization rather than on every loop iteration. Bug fixes address deterministic CSR ordering, privilege mode handling in virtualized environments, and stack allocation for algorithm tests, while new documentation with Mermaid flowcharts and API details should help developers understand and debug runtime behavior more easily.
Links: Repository · v1.7.0
\nofficial · added 2025-12-03
", "summary": "The loader and scheduler in RiescueD have been refactored with explicit interfaces that make their control flow clearer and reduce redundant checks—`test_setup` now runs once during scheduler initialization rather than on every loop iteration. Bug fixes address deterministic CSR ordering, privilege mode handling in virtualized environments, and stack allocation for algorithm tests, while new documentation with Mermaid flowcharts and API details should help developers understand and debug runtime behavior more easily.", "date_published": "2025-12-03T19:29:44Z", "tags": [ "official", "release" ] }, { "id": "https://github.com/tenstorrent/tt-torch/releases/tag/0.4.0", "url": "https://github.com/tenstorrent/tt-torch/releases/tag/0.4.0", "title": "tt-torch 0.4.0", "content_html": "tt-torch released 0.4.0. Repository: https://github.com/tenstorrent/tt-torch
\n\nLinks: Repository · 0.4.0
\nofficial · added 2025-09-29
", "summary": "tt-torch released 0.4.0. Repository: https://github.com/tenstorrent/tt-torch", "date_published": "2025-09-29T22:23:47Z", "tags": [ "official", "release" ] }, { "id": "https://github.com/tenstorrent/polaris/releases/tag/pre_perfmodel_merge", "url": "https://github.com/tenstorrent/polaris/releases/tag/pre_perfmodel_merge", "title": "polaris pre_perfmodel_merge", "content_html": "polaris released pre_perfmodel_merge. Repository: https://github.com/tenstorrent/polaris
\n\nLinks: Repository · pre_perfmodel_merge
\nofficial · added 2025-09-19
", "summary": "polaris released pre_perfmodel_merge. Repository: https://github.com/tenstorrent/polaris", "date_published": "2025-09-19T18:16:27Z", "tags": [ "official", "release" ] }, { "id": "https://github.com/tenstorrent/riscv_arch_tests/releases/tag/v0.2.0%2Baligned-access", "url": "https://github.com/tenstorrent/riscv_arch_tests/releases/tag/v0.2.0%2Baligned-access", "title": "riscv_arch_tests v0.2.0+aligned-access", "content_html": "This release brings byte-alignment compliance testing to the RISC-V architecture test suite, ensuring that RV IMFA tests across all privilege levels (Machine, Supervisor, and User) and configurations (bare metal and paging disabled) properly enforce alignment constraints. The CI pipeline now validates that these tests correctly reject misaligned memory accesses, catching potential issues where implementations might incorrectly permit unaligned data or instruction fetches that the spec forbids.
\n\nLinks: Repository · v0.2.0+aligned-access
\nofficial · added 2025-01-23
", "summary": "This release brings byte-alignment compliance testing to the RISC-V architecture test suite, ensuring that RV IMFA tests across all privilege levels (Machine, Supervisor, and User) and configurations (bare metal and paging disabled) properly enforce alignment constraints. The CI pipeline now validates that these tests correctly reject misaligned memory accesses, catching potential issues where implementations might incorrectly permit unaligned data or instruction fetches that the spec forbids.", "date_published": "2025-01-23T17:16:16Z", "tags": [ "official", "release" ] }, { "id": "https://github.com/tenstorrent/tt-buda/releases/tag/v0.19.3", "url": "https://github.com/tenstorrent/tt-buda/releases/tag/v0.19.3", "title": "tt-buda v0.19.3", "content_html": "This release brings support for ten additional model variants—including Phi2, Qwen1.5-0.5B, and YOLOX—along with initial Wormhole n300 dual-chip support for the TT-LoudBox and TT-QuietBox systems. Performance improvements are mixed across architectures: Grayskull sees solid gains on HRNet (41%), while Wormhole n300 single-chip shows 14% uplift on Falcon-7B, though some models like FLAN-T5 have regressed. CNN models now work on 4-chip and 8-chip MMIO configurations, and the compiler received stability fixes alongside improved documentation.
\n\nLinks: Repository · v0.19.3
\nofficial · added 2024-09-24
", "summary": "This release brings support for ten additional model variants—including Phi2, Qwen1.5-0.5B, and YOLOX—along with initial Wormhole n300 dual-chip support for the TT-LoudBox and TT-QuietBox systems. Performance improvements are mixed across architectures: Grayskull sees solid gains on HRNet (41%), while Wormhole n300 single-chip shows 14% uplift on Falcon-7B, though some models like FLAN-T5 have regressed. CNN models now work on 4-chip and 8-chip MMIO configurations, and the compiler received stability fixes alongside improved documentation.", "date_published": "2024-09-24T21:01:08Z", "tags": [ "official", "release" ] }, { "id": "https://github.com/zk4x/zyx/releases/tag/v0.14.0", "url": "https://github.com/zk4x/zyx/releases/tag/v0.14.0", "title": "zyx v0.14.0", "content_html": "zyx released v0.14.0. Repository: https://github.com/zk4x/zyx
\n\nLinks: Repository · v0.14.0
\ncommunity · added 2024-09-22
", "summary": "zyx released v0.14.0. Repository: https://github.com/zk4x/zyx", "date_published": "2024-09-22T13:54:32Z", "tags": [ "community", "release" ] } ] }