Note:

The performance figures above represent the theoretical peak for a single GPU, focusing on AI and deep learning workloads. The values are approximate and depend on the specific data format and whether sparsity is utilized. For instance, the A100 performance is based on FP16/BF16 precision with sparsity. The H100 and B200 performance figures are based on the newer FP8 and FP4 data formats with sparsity, respectively, which are optimized for AI inference and training. The significant increase in performance for FY2024 with the H100 reflects its widespread adoption. The Blackwell B200 GPU, announced in March 2024, offers a substantial leap in performance. This rapid increase in computational power has been a key driver of Nvidia's datacenter revenue growth.