Hardware Detection
Detect and verify your Tenstorrent hardware using the tt-smi command-line tool.
Before You Begin: Is Your System Already Set Up?
If you're using:
- ✅ Tenstorrent Cloud - Your environment is pre-configured
- ✅ Quietbox with preinstalled image - tt-smi and drivers are already installed
- ✅ Managed system - Your sysadmin likely ran tt-installer already
You can skip directly to running tt-smi below!
Need to Install? Use tt-installer 2.0
If tt-smi is not found on your system, the fastest way to set up is with tt-installer 2.0 (recommended for 95% of users):
# One-command installation (5-15 minutes)
/bin/bash -c "$(curl -fsSL https://github.com/tenstorrent/tt-installer/releases/latest/download/install.sh)"
What it installs:
- ✅ tt-smi (this tool)
- ✅ Kernel drivers
- ✅ Firmware updates
- ✅ HugePages configuration
- ✅ tt-metalium containers (1GB standard, optional 10GB with demos)
- ✅ Python environment
Consult the tt-installer repo for more information.
What This Does
The tt-smi command scans your system for connected Tenstorrent devices and displays their status, including:
- Device model and ID
- PCIe information
- Temperature and power status
- Driver version
Run the Command
This command will scan for connected Tenstorrent devices:
tt-smi
tt-smi
Understanding Your Hardware
After running tt-smi, you'll see information about your Tenstorrent hardware. To get structured output, use:
tt-smi -s
Example output (JSON format):
{
"board_info": {
"board_type": "n150",
"coords": "0,0"
},
"telemetry": {
"voltage": 0.8,
"current": 25.3,
"power": 20.24,
"temperature": 45.0
}
}
Hardware Types
Look for the board_type field to identify your hardware:
Wormhole Family (2nd Generation):
n150 - Single chip, 72 Tensix cores
- Best for: Development, prototyping, single-user workloads
- Context limit: 64K tokens for most models
- Common in: Cloud instances, development workstations
n300 - Dual chip (2 chips), 144 Tensix cores
- Best for: Higher throughput, longer context, multi-user serving
- Context limit: 128K tokens
- Tensor parallelism: TP=2 (uses both chips)
t3k - Eight chips (8 chips), 576 Tensix cores
- Best for: Large models (70B+), production serving
- Context limit: 128K tokens
- Tensor parallelism: TP=8 (uses all chips)
Blackhole Family (Latest Generation):
p100 - Single chip (cloud/standalone deployments)
- Best for: Similar to N150 but with newer architecture
- Context limit: 64K tokens
- Status: Some models validated, others experimental
p150 - Dual chip (higher performance)
- Best for: Similar to N300 but with improvements
- Context limit: 128K tokens
- Status: Check official documentation for validated configurations
p300/p300c - Single chip (QuietBox variant)
- Architecture: Blackhole (identical to P100)
- Common in: Multi-device QuietBox Tower systems
- MESH_DEVICE: Use P100 for single-chip lessons
- Example: 4x P300c = 4 separate single-chip devices
Blackhole Architecture Equivalence: All Blackhole cards (P100, P150, P300/P300c) share the same instruction set and capabilities. Lessons supporting P100 will work on P300/P300c without modification.
QuietBox Multi-Device Detection:
If you have a QuietBox Tower (4x P300c), tt-smi will show 4 devices:
Device 0: 0000:01:00.0 | P300c | FW 19.4.0.0
Device 1: 0000:02:00.0 | P300c | FW 19.4.0.0
Device 2: 0000:03:00.0 | P300c | FW 19.4.0.0
Device 3: 0000:04:00.0 | P300c | FW 19.4.0.0
Each device is a separate single-chip Blackhole card. For single-chip lessons, use device 0. For multi-chip lessons, all 4 devices are available for workload distribution.
Quick Hardware Check
Extract just your hardware type:
tt-smi -s | grep -o '"board_type": "[^"]*"'
Output: "board_type": "n150"
Save this for later lessons! You'll need to know your hardware type when configuring models and servers.
Expected Output
When you run tt-smi, you should see:
For Wormhole hardware (N150, N300, T3K):
Device 0: Wormhole
Board Type: n150
PCIe: Bus 0x01, Device 0x00
Firmware Version: 1.2.3
Temperature: 45°C
Power: 20W
For Blackhole hardware (P100, P150):
Device 0: Blackhole
Board Type: p100
PCIe: Bus 0x01, Device 0x00
Firmware Version: 2.0.1
Temperature: 42°C
Power: 18W
Multiple devices:
Device 0: Wormhole (n150)
Device 1: Wormhole (n150)
Device 2: Wormhole (n150)
Device 3: Wormhole (n150)
Troubleshooting: No Hardware Detected
Don't worry if tt-smi doesn't detect your hardware immediately. You can usually fix this easily. Try these steps:
Check 1: Hardware Connection
Verify card is detected by PCIe:
lspci | grep -i tenstorrent
Expected output:
01:00.0 Processing accelerators: Tenstorrent Inc. Device [model]
If nothing appears:
- Card isn't properly seated
- Power cable not connected
- System needs to be restarted
- BIOS/UEFI settings may need adjustment
Check 2: Driver Installation
Check if tt-smi is installed:
which tt-smi
Expected: /usr/local/bin/tt-smi or similar path
If not found:
- Driver not installed
- Install from: tt-smi installation guide
Check 3: Permissions
Try with sudo:
sudo tt-smi
If this works but tt-smi alone doesn't:
- Permissions issue
- Add your user to the appropriate group
- Or use sudo for now
Fix permissions (Linux):
sudo usermod -a -G tenstorrent $USER
# Log out and back in for group changes to take effect
Check 4: Device Reset
If device appears but shows errors:
tt-smi -r
What this does:
- Resets the Tenstorrent device
- Clears any error states
- Takes ~5-10 seconds
If reset fails:
# Full state cleanup (requires sudo)
sudo pkill -9 -f tt-metal
sudo pkill -9 -f vllm
sudo rm -rf /dev/shm/tenstorrent* /dev/shm/tt_*
tt-smi -r
Still Having Issues?
Check system logs:
dmesg | grep -i tenstorrent
Get help:
Common issues:
- "No kernel driver loaded" → Reinstall tt-smi/drivers
- "Device initialization timeout" → Try device reset (
tt-smi -r) - "Permission denied" → Add user to tenstorrent group or use sudo
QuietBox Multi-Device Systems
What is QuietBox? QuietBox is a Tenstorrent multi-chip development system. The QuietBox Blackhole Tower contains 4x P300c cards (4 separate single-chip Blackhole devices).
Key Concepts:
4x P300c ≠ 4-chip System
- 4x P300c = 4 separate cards, each with 1 Blackhole chip
- Total: 4 devices, each independently addressable
- Each device runs in P100 mode (single Blackhole chip)
Device Enumeration:
tt-smi -s # Shows all 4 devices
{
"device_0": {"board_type": "p300c", "pci_bus": "0000:01:00.0"},
"device_1": {"board_type": "p300c", "pci_bus": "0000:02:00.0"},
"device_2": {"board_type": "p300c", "pci_bus": "0000:03:00.0"},
"device_3": {"board_type": "p300c", "pci_bus": "0000:04:00.0"}
}
Configuration for Lessons:
Single-Chip Lessons (Most lessons):
- Use device 0 only:
export TT_METAL_DEVICE_ID=0 - Set MESH_DEVICE:
export MESH_DEVICE=P100 - Architecture:
export TT_METAL_ARCH_NAME=blackhole
Multi-Device Lessons (Advanced):
- Use all devices:
TT_METAL_NUM_DEVICES=4 - See Lesson 15 (Metalium Cookbook - Particle Life) for multi-device example
- Achieves 2x speedup on 4x P300c through workload parallelization
Troubleshooting:
- If script says "Unknown board type 'P300C'": Treat as P100 (single Blackhole)
- Multi-chip mesh initialization: All 4 devices will initialize fabric
- Device reset: Use
tt-smi -rcarefully (close all processes first)
What You Learned
- ✅ How to detect Tenstorrent hardware with
tt-smi - ✅ Understanding different hardware types (N150, N300, T3K, P100, P150, P300/P300c)
- ✅ Using
tt-smi -sfor structured JSON output - ✅ Identifying your specific hardware for later lessons
- ✅ QuietBox multi-device system configuration
- ✅ Troubleshooting hardware detection issues
Next step: Now that you know your hardware, verify your tt-metal installation works correctly.
Continue to Lesson 2: Verify Installation!
Learn More
- tt-smi documentation: github.com/tenstorrent/tt-smi
- Hardware specs: tenstorrent.com/hardware
- Community support: discord.gg/tenstorrent