NVIDIA GB200 Grace Blackwell Superchip · Blackwell

BF16, dense

5,000 TFLOPS FP16/BF16 Tensor Core2 360 PFLOPS 10 PFLOPS NVIDIA GB200 NVL72 (10 PFLOPS on the page; read from a table with a column for each product, whose head reads “GB200 Grace Blackwell Superchip Configuration 36 Grace CPU | 72 Blackwell GPUs 1 Grace CPU | 2 Blackwell GPU”; halved, because the page gives it with sparsity)

FP8, dense

10,000 TFLOPS FP8/FP6 Tensor Core2 720 PFLOPS 20 PFLOPS NVIDIA GB200 NVL72 (20 PFLOPS on the page; read from a table with a column for each product, whose head reads “GB200 Grace Blackwell Superchip Configuration 36 Grace CPU | 72 Blackwell GPUs 1 Grace CPU | 2 Blackwell GPU”; halved, because the page gives it with sparsity)

FP4, dense

20,000 TFLOPS NVFP4 Tensor Core2 1,440 | 720 PFLOPS 40 | 20 PFLOPS NVIDIA GB200 NVL72 (20 PFLOPS on the page; read from a table with a column for each product, whose head reads “GB200 Grace Blackwell Superchip Configuration 36 Grace CPU | 72 Blackwell GPUs 1 Grace CPU | 2 Blackwell GPU”)

Memory

372 GB GPU Memory | Bandwidth 13.4 TB HBM3E | 576 TB/s 372 GB HBM3E | 16 TB/s NVIDIA GB200 NVL72 (read from a table with a column for each product, whose head reads “GB200 Grace Blackwell Superchip Configuration 36 Grace CPU | 72 Blackwell GPUs 1 Grace CPU | 2 Blackwell GPU”)

Memory bandwidth

16 TB/s GPU Memory | Bandwidth 13.4 TB HBM3E | 576 TB/s 372 GB HBM3E | 16 TB/s NVIDIA GB200 NVL72 (read from a table with a column for each product, whose head reads “GB200 Grace Blackwell Superchip Configuration 36 Grace CPU | 72 Blackwell GPUs 1 Grace CPU | 2 Blackwell GPU”)

Figures are per superchip (1 Grace CPU and 2 Blackwell GPUs). The page gives FP4 as sparse | dense (dense used) and gives BF16 and FP8 only as sparse. No readable NVIDIA page gives a power figure for one GB200 superchip (the Blackwell datasheet gives up to 1,200 W per GPU in GB200 NVL72), so power is left out. The 2025-02-26 earnings release already reported Blackwell sales; the 2025-05-28 release is the first that names NVL72 in full production.

1. 18 Mar 2024 Promised Blackwell-based products available from partners starting later this year in 2024. Blackwell-based products will be available from partners starting later this year. NVIDIA Blackwell Platform Arrives to Power a New Era of Computing
2. 28 May 2025 Shipped: Blackwell NVL72 in full-scale production across system makers and cloud providers. Our breakthrough Blackwell NVL72 AI supercomputer - a 'thinking machine' designed for reasoning- is now in full-scale production across system makers and cloud service providers NVIDIA Announces Financial Results for First Quarter Fiscal 2026

Epoch AI’s names for this chip: NVIDIA GB200.

• Tesla V100 SXM2 417 gigaflops of dense FP16 compute per watt, at 300 W
• A100 SXM 80GB 780 gigaflops of dense BF16 compute per watt, at 400 W
• H100 SXM 1,414 gigaflops of dense BF16 compute per watt, at 700 W
• H200 SXM 1,414 gigaflops of dense BF16 compute per watt, at 700 W
• GH200 Grace Hopper Superchip 990 gigaflops of dense BF16 compute per watt, at 1,000 W
• B200 2,200 gigaflops of dense BF16 compute per watt, at 1,000 W
• B300 (Blackwell Ultra) 2,000 gigaflops of dense BF16 compute per watt, at 1,100 W
• Rubin GPU 4,000 teraflops of dense BF16 compute a chip; NVIDIA publishes no power figure