Kernel Leaderboard (www.gpumode.com)

The Kernel Leaderboard shows active community-driven benchmarks for low-level GPU kernels across different architectures and workloads. Recent top results include amd-gemm-rs on MI300x8 with pank2025 winning at 530.683 μs (closely followed at 532.916 μs and 534.345 μs), amd-all2all also on MI300x8 led by fanwenjie at 547.474 μs, and a trimul test spanning A100/B200/H100/MI300 where Arseni Ivanov tops at 1800.041 μs. The amd-mixture-of-experts challenge on MI300 has concluded with ColorsWind first at 6654.929 μs, while a grayscale_py_b200-dev benchmark on B200 is being led by charles_irl at 688.308 μs. Several contests remain open (time left shown), indicating ongoing tuning and submissions. For the AI/ML community this is a practical, real-world snapshot of kernel-level performance and cross-vendor competitiveness. Close margins in the fastest GEMM/all-to-all times highlight micro-optimization opportunities (memory layout, vectorization, synchronization) that can translate to meaningful model throughput gains at scale. The variety of hardware (MI300, A100, H100, B200) and workloads (GEMM, all2all, mixture-of-experts, image grayscale) also helps teams choose architectures and prioritize kernel work for latency- or throughput-sensitive inference and training. Continued public leaderboards foster reproducibility, rapid iteration, and community sharing of best practices for kernel tuning.