Blackwell SM microarchitecture。SM(Streaming Multiprocessor) 170 core/SM・Tensor Memory(TMem 256KB/SM・low-precision optimized)・Thread Block Cluster・5th Gen Tensor Core(FP4 double throughput)・RT Core 4th Gen(1.75x Ray Intersection)・NVENC 10th Gen(AV1 + H.266)・NVDEC 9th Gen・L1/Shared 228KB/SM・L2 96MB total(RTX 5090)・GB202 170 SM=21760 core、2026年CUDA 13 programming model拡張。