Startup claims to beat cuDNN
A startup called Standard Kernel reportedly raised $20M (with Jeff Dean/CoreWeave backing) and claims its GPU‑software can deliver 0.8×–4× speedups vs. NVIDIA's cuDNN on H100s, per social posts reported. Those performance claims are already prompting discussion about third‑party GPU software versus vendor‑supplied stacks.
The round was led by Jump Capital and announced from Palo Alto on March 11, 2026, with participation from General Catalyst, Felicis, Cowboy Ventures, Link Ventures, Essence VC, and Ericsson Ventures. (prnewswire.com) Standard Kernel describes its product as an AI-driven system that generates hardware‑specialized, instruction‑level kernels intended to replace one‑size‑fits‑all libraries and produce “day‑one” peak performance on new hardware. (prnewswire.com) The company has published a technical note on high‑fidelity kernel benchmarking that calls out concrete pitfalls — isolating GPU resources, warming up devices, synchronizing correctly, and using low‑overhead timers for repeatable measurements. (standardkernel.com) Academic and open‑source benchmarks remain cautious: the KernelBench / OpenReview evaluation found generated kernels matched the PyTorch baseline in under 20% of cases on their suite, and the hardest task classes showed pass rates well below 25%. (openreview.net) Follow‑on projects and papers (for example, robust‑kbench and KernelBench on GitHub) are explicitly focused on verification, reproducibility, and correctness checks for automated kernel engineering. (arxiv.org) Vendor stacks continue to evolve at scale: a recent large‑scale MLPerf training submission reported training a 175B‑parameter model in 3.4 minutes on a multi‑thousand‑GPU cluster, underscoring the ongoing software optimization arms race. (developer.nvidia.com) The company says the new capital will fund product engineering and commercial partnerships as it moves from partner tests toward broader customer integrations; the press release frames this as scaling the kernel‑generation platform and expanding partner evaluations. (prnewswire.com)
