Hardware Benchmarks

Daniel discusses the limitations of current hardware benchmarks, highlighting the lack of real-world applicability and the overly specific tuning of hardware backends. He questions how benchmarks like ML Perf could be improved to provide more indicative results.