Autotuning

In episode 72, AMD’s Anush Elangovan describes an older pattern: define a search space, sample candidate configurations, and measure the results. An AI-guided tuner can make more informed choices about what to test next. The important distinction for a builder is what gets measured and whether a faster candidate still produces correct output.

AMD describes Hyperloom as an inference optimization loop that profiles a workload, searches serving, framework, and GPU-kernel changes, then remeasures candidates end to end. It keeps a change only when throughput improves and accuracy holds against the baseline. AMD reports that Hyperloom has optimized more than 14,000 models. In episode 72, Anush says he thinks one case optimized about 14,000 models in one pass; AMD’s page doesn’t confirm or contradict that detail.