AI teams waste time manually testing new hardware against their workloads.
Build a cloud service that runs standardized ML benchmarks across different accelerators. Offer side-by-side comparisons of throughput, memory usage, and power efficiency.
Cloud providers and hardware vendors would pay to showcase their platforms' performance. Charge for custom benchmarking against proprietary models.
Start with basic image and text model benchmarks on a few GPU types.
Risk is hardware vendors creating their own benchmark suites.