TensixFuse

Fused and block-float ML kernels for Tensix cores, written in TT-Lang and by hand in TT-Metalium C++, measured on Tenstorrent's simulators. Quantifies how much DRAM traffic kernel fusion and data reuse save on a fused matmul-bias-ReLU, and how far GPT-2 MLP weights can be pushed into bfloat8_b and bfloat4_b before accuracy drops, with an analytic traffic model that matches the simulator tile for tile. CI runs the studies on simulated Wormhole and Blackhole on every change.