Comparison of NVIDIA Data Center Chips

H100, H200, GH200, B200, and GB200 at a glance

Feature H100 (SXM) H200 (GCP Typical) GH200 (Superchip) B200 (Standalone) GB200 (Superchip)
Architecture Hopper Hopper (Upgraded) Grace Hopper Blackwell Grace Blackwell
FP16 Tensor Core (Sparse) 989 TFlops 989 TFlops 989 TFlops 2,250 TFlops 5,000 TFlops (Total)¹
FP8 Tensor Core (Sparse) 3,958 TFlops 3,958 TFlops 3,958 TFlops 4,500 TFlops 9,000 TFlops (Total)¹
FP4 Tensor Core (Sparse)² 9,000 TFlops 18,000 TFlops (Total)¹
GPU Memory (VRAM) 80 GB HBM3 141 GB HBM3e 96 GB or 144 GB HBM3e 192 GB HBM3e 384 GB HBM3e (2× 192 GB)
CPU Model External (x86) External (x86) NVIDIA Grace External (x86) NVIDIA Grace
CPU Memory (RAM) Varies (e.g. 2 TB) Varies (e.g. 2 TB) 480 GB (LPDDR5X) Varies (e.g. 2 TB+) 480 GB (LPDDR5X)
CPU-to-GPU Link PCIe Gen5 PCIe Gen5 NVLink-C2C PCIe Gen5 NVLink-C2C

¹ GB200 combines two B200 GPUs with a Grace CPU; "Total" figures are the combined throughput for the dual-GPU superchip.

² FP4 (4-bit floating point) is introduced with the Blackwell architecture. Hopper-based chips (H100, H200, GH200) do not have native FP4 Tensor Core support. FP8 was introduced with Hopper and is supported by all chips in this table.