NVIDIA Blackwell 5th Gen Tensor Core。FP4(Float 4-bit・new Blackwell)・FP6/FP8 Hopper継続・FP4 throughput 2x FP8(同電力)・Inference特化(Llama 3.3 70B FP4 + GB200で 1ms/token)・Multi Frame Generation学習FP4活用・Sparsity 2:4 Pattern(2 of 4 zero・throughput 2x effective)・MicroScaling(MX) format(FP6/FP8 hybrid・OCP standard)・PyTorch FP4 native対応(2025年Q3-)・Hopper FP8 vs Blackwell FP4性能比 +60%・2026年 RTX PRO 6000 Blackwell + FP4 Inference推論主流化。