Run GPU kernel tuning as an overnight search loop
Profile the model, find the kernel or fusion hotspots, and treat later kernel changes as candidates in a search. Run the trials in parallel on idle GPUs, including overnight, and keep going while there is compute until the workload is performant. Anush Elangovan, Corporate VP of AI Software and Solutions at AMD, laid this out on Chain of Thought in September 2026. ROCm 10 is a stack an agent can…
Anush Elangovan, AMD's Corporate VP of AI Software and Solutions, advises running GPU kernel tuning as an overnight search loop. After profiling the model, GEAK is used to search for better utilization of slow kernels while maintaining numerical accuracy. The profiling and analysis step identifies kernel hotspots or fusion hotspots.
The GEAK toolkit searches for better optimization while preserving the same numerical accuracy. A tolerance is set before the first trial to ensure that any config breaking it is discarded. The speed of the kernel is prioritized only for configs that still match the reference. The tuning loop runs until there is compute left, using idle GPUs.
Hyperloom can optimize on the fly while you sleep, and parallel runs can be performed when the system is not in use. The tuner structures the job into well-defined problems and goals, with inputs, reference outputs, tolerance, the metric being optimized, and a stop rule. The process involves running the kernel, comparing against a reference with a max absolute or relative difference, timing, and repeating the process. The goal is to find the best configuration within the GPU hours budget.
Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
