Operator Fusion for LLM Inference on the Tensix Architecture

In three linesOptimization study for LLM inference on Tenstorrent's Tensix architecture. Operator fusion (RMSNorm + matrix multiplication) reduces DRAM accesses and latency: -37.44% attention, -15.89% MLP on Qwen2.5-0.5B, Qwen3-0.6B/4B. NoC-based multicast for multi-core parallelism.

Summary generated by Claude — human-verified