GH200 is NVIDIA’s “Grace Hopper Superchip” which combines a single Grace CPU with a H100 or H200 GPU.1
The GH200 platform also supports 18x NVLink 4 lanes out of each Hopper GPU, allowing for two GH200 superchips to be coherently connected (forming “GH200 NVL”) or up to 256 GH200 superchips using NVLink Switch.1 NVIDIA calls this DGX GH200.
Performance
The following are theoretical maximum performance in TFLOPS.2 They appear identical to H100.
| Data Type | VFMA | Matrix | Sparse |
|---|---|---|---|
| FP64 | 34 | 67 | |
| FP32 | 67 | ||
| TF32 | 494 | ||
| FP16 | 990 | ||
| BF16 | 990 | ||
| FP8 | 1979 | ||
| INT32 | 1979 | ||
| INT8 | 1979 |
I haven’t read it, but Torsten Hoefler’s group published a paper on Alps that contains benchmark results for GH200: “Understanding Data Movement in Tightly Coupled Heterogeneous Systems: A Case Study with the Grace Hopper Superchip”