Chips · accelerators · NVIDIA
NVIDIA GH200 Grace Hopper Superchip · Hopper
990 gigaflops of dense BF16 compute per watt, at 1,000 W
Shipped 28 May 2023; shipped by the first date promised. Epoch AI names it at no AI data center and 10 GPU clusters.
Embed
The frame credits NVIDIA, Epoch AI, AI data centers and Epoch AI, Data on GPU clusters and links back to its page. Preview it.
Cite
Each figure in NVIDIA’s own words
one chip, dense compute; memory is high-bandwidth memory- BF16, dense
- 990 TFLOPS
BFLOAT16 Tensor Core 1,979 teraFLOPS* | 990 teraFLOPS
NVIDIA GH200 Grace Hopper Superchip datasheet (Mar 2025) - FP8, dense
- 1,979 TFLOPS
FP8 Tensor Core 3,958 teraFLOPS* | 1,979 teraFLOPS
NVIDIA GH200 Grace Hopper Superchip datasheet (Mar 2025) - Memory
- 96 GB
High-bandwidth memory (HBM) size 96GB HBM3 | 144GB HBM3e
NVIDIA GH200 Grace Hopper Superchip datasheet (Mar 2025) - Memory bandwidth
- 4 TB/s
Memory bandwidth Up to 4TB/s | Up to 4.9TB/s
NVIDIA GH200 Grace Hopper Superchip datasheet (Mar 2025) - Maximum power
- 1,000 W
Configurable 450 to 1000W (Memory + CPU + GPU)
NVIDIA GH200 Grace Hopper Superchip datasheet (Mar 2025) - Compute per watt
- 990 GFLOPS/Wdense BF16 over maximum power; 1,979 GFLOPS/W at FP8
Figures are for the first GH200 with 96GB HBM3 (4 TB/s), which matches the 2023 production date; the later HBM3e version has 144GB and 4.9 TB/s. The datasheet gives sparse and dense; dense is used. Power is for the whole module (memory, CPU and GPU).
2 dated promises and the shipment
each date links to the page it is quoted from- 22 Mar 2022 Promised Grace Hopper Superchip expected to be available in the first half of 2023 in the first half of 2023.
Both it and the Grace Hopper Superchip are expected to be available in the first half of 2023.
NVIDIA Introduces Grace CPU Superchip - 28 May 2023 Promised systems with GH200 expected to be available beginning later this year in 2023.
Systems with GH200 Grace Hopper Superchips are expected to be available beginning later this year.
NVIDIA Grace Hopper Superchips Designed for Accelerated Generative AI Enter Full Production - 28 May 2023 Shipped: GH200 Grace Hopper Superchip in full production.
today announced that the NVIDIA® GH200 Grace Hopper Superchip is in full production
NVIDIA Grace Hopper Superchips Designed for Accelerated Generative AI Enter Full Production
Named at no AI data center Epoch AI tracks
H100-equivalents and power are the whole site’s today, with every chip it holds53,508 chips in 10 GPU clusters
Epoch AI’s list of AI supercomputers, which stops in July 2025| Cluster | Owner | Country | Chips | First operational |
|---|---|---|---|---|
| Jupiter, Jülich | EuroHPC JU,Julich Supercomputing Center | Germany | 23,536 | 2025-06-10 |
| Alps Supercomputer Phase 2 | ETH Domain | Switzerland | 10,752 | 2024-09-17 |
| University of Bristol Isambard-AI | UK Research and Innovation | United Kingdom | 5,448 | 2025-06-15 |
| ExxonMobil Discovery 6 | ExxonMobil | United States of America | 4,032 | 2025-06-15 |
| CEA EXA1-HE Phase 3 | French Alternative Energies and Atomic Energy Commission (CEA) | France | 2,688 | 2025-06-15 |
| Los Alamos NL Venado | US Department of Energy | United States of America | 2,560 | 2024-04-16 |
| CEA EXA1-HE Phase 2 | French Alternative Energies and Atomic Energy Commission (CEA) | France | 1,908 | 2024-04-17 |
| JCAHPC Miyabi | Joint Center for Advanced High Performance Computing (JCAHPC) | Japan | 1,120 | 2025-01-15 |
| NVIDIA Helios | Nvidia | Not given | 1,024 | 2023-11-15 |
| AGH Cyfronet Helios | AGH University of Krakow | Poland | 440 | 2024-06-01 |
Epoch AI’s names for this chip: NVIDIA GH200.
Other NVIDIA chips
- 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
- B200 2,200 gigaflops of dense BF16 compute per watt, at 1,000 W
- GB200 Grace Blackwell Superchip 5,000 teraflops of dense BF16 compute a chip; NVIDIA publishes no power figure
- 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
- NVIDIA’s own pages, each linked beside the figure or date it gives. Each page was read and each quote checked on 23 Oct 2026.
- Where it runs: Epoch AI, AI data centers and Epoch AI, GPU clusters, CC BY 4.0.