Modular beat Nvidia's cuBLAS kernels on B200s in 170 LOC (twitter.com)

Modular announced that a compact 170-line kernel implementation outperformed Nvidia’s cuBLAS kernels on B200 accelerators. The result—reported as a head-to-head benchmark on B200 hardware—shows that a small, hand-tuned or compiler-assisted kernel can beat a vendor-optimized library in real workloads, challenging assumptions about how much code and complexity are required to reach peak performance on new GPUs/accelerators. Technically, the win highlights the value of focused kernel design: careful tiling, memory-hierarchy utilization, vectorization, and warp/thread scheduling (and likely light-weight auto-tuning) can unlock higher utilization on unfamiliar hardware than a general-purpose library tuned for older architectures. For the AI/ML community this matters because faster, simpler kernels reduce development overhead, enable rapid experimentation for LLM training and inference, and weaken vendor lock-in—encouraging open tooling (Triton/TVM/MLIR-style stacks) and community-driven optimization. In short, the demonstration underscores that lean, portable kernel engineering can compete with entrenched libraries on new accelerators, accelerating adoption of alternative hardware and driving improvements in compiler and kernel-generation toolchains.