Infatoshi/kernelbench-v3-problems
收藏资源简介:
KernelBench-v3是一个GPU内核生成基准测试数据集,涵盖了10种前沿模型在3种NVIDIA GPU(RTX 3090、H100、B200)上的表现,每种GPU包含43至58个问题。数据集包含多个级别的问题定义,从简单操作到复杂架构块和新颖层,以及图形问题和特定硬件优化问题。数据集采用速度提升作为评分标准,要求解决方案在通过精度验证后才能计算速度提升。数据集还提供了参考实现和测试输入,便于用户进行基准测试和性能评估。
KernelBench-v3 is a GPU kernel generation benchmark dataset covering 10 frontier models across 3 NVIDIA GPUs (RTX 3090, H100, B200), with 43–58 problems per GPU. The dataset includes multiple levels of problem definitions, from simple operations to complex architectural blocks and novel layers, as well as graphics problems and hardware-specific optimizations. The dataset uses speedup-over-baseline as the scoring metric, requiring solutions to pass precision validation before calculating speedup. It also provides reference implementations and test inputs for benchmarking and performance evaluation.




