Megadose The AI curve, compressed.

WaferTetris: Efficient Freeform Mapping on Wafer-Scale Accelerators

· ArXiv · AI/CL/LG ·
WaferTetris delivers up to 4.42× the LLM decoding throughput of an automatic template-based baseline on the same wafer fabric size.

The system packs computation into freeform layouts, improving processor and memory utilization while preserving existing task code. The authors evaluated it on Cerebras WSE-3 using LLM decoding and scientific stencil benchmarks. Against the manually optimized WaferLLM, it supports an 80,000-token context for LLaMA3.1-8B, versus 44,000 tokens. At matched context lengths, per-user throughput equals or exceeds WaferLLM’s, reaching up to 1.10×.

ArXiv · AI/CL/LG's note

score 5

Categories: Research