AMD: overnight AI GPU kernel tuning optimized 14,000 models

AMD's corporate VP of AI software Anush Elangovan described a GPU kernel tuning loop: profile hotspots, search kernel variants while enforcing numerical tolerance, and run trials in parallel on idle GPUs overnight. He said Hyperloom optimized about 14,000 models in one pass, with GEAK searching for faster kernels.

• Hyperloom optimized about 14,000 models in one pass
• GEAK searches faster kernels while keeping numerical accuracy
• Tolerance is fixed before the first trial and rejects configs
• Agents can install ROCm 10 and use it to serve a model