We now have a benchmark that turns GPU performance into a "speed-of-light" score so you can see how much hardware headroom is left. Excited to see which tools top the leaderboard. — NVIDIA AI Developer (@NVIDIAAIDev) How close can you get to the speed of light? ⚡ Introducing SOL-ExecBench from NVIDIA — a benchmark for real-world GPU kernels that measures performance against hardware-grounded Speed-of-Light (SOL) bounds, not just software baselines. It includes 235 CUDA kernel optimization problems extracted from 124 production and emerging AI models, spanning forward and backward workloads across BF16, FP8, and NVFP4 on NVIDIA Blackwell GPUs. Dive in:
🏆 Leaderboard: research.nvidia.com/benchmarks/sol-execbench
🤗 Dataset: huggingface.co/datasets/nvidia/SOL-ExecBench
💻 Evaluator: github.com/nvidia/sol-execbench
📑 Paper: arxiv.org/abs/2603.19173 [Translated from EN to English]

→ View original post on X — @nvidiaai, 2026-03-20 20:22 UTC