Still 535.xx minimum, but 550+ recommended for Blackwell features.
• Lower kernel launch overhead (big for H100/H200) • Official Blackwell support • cuBLAS/cuDNN FP8/16 perf wins • Drops Kepler/Maxwell cuda 12.6 news
CUDA 12.6 just dropped ⚡
If you’re running LLM inference, large-scale simulations, or building for Blackwell – yes . For older data center GPUs (V100, A100), test first but the improvements are solid. Still 535