NVIDIA · Datacenter System
DGX GH200 (256× Grace Hopper)
Grace Hopper
Prefill · prompt processing
Not benchmarked
Inference · generation speed
Not benchmarked
GPU / CPU chip
GH200 Grace Hopper Superchip / 256× NVIDIA Grace
RAM / VRAM
147,456 GB coherent
122,880 GB system RAM
No verified model-specific throughput measurement is published here yet. Prefill and generation rates depend on the model, precision, batch size and runtime.
Hardware specifications
| Product family | DGX |
| Model / configuration | DGX GH200 |
| Released | 2023 |
| Manufacturer | NVIDIA |
| Hardware category | Datacenter System |
| Architecture | Grace Hopper |
| Form factor | Supercomputer scale-out system |
| CPU | 256× NVIDIA Grace |
| Accelerator model | GH200 Grace Hopper Superchip |
| Accelerator count | 256 |
| Interconnect | NVLink Switch System |
| Interconnect bandwidth | 900 GB/s |
Where to buy
No verified retailer listings have been added for this hardware yet.
Retailer links do not guarantee stock or a current price.
Observed prices
No source-backed price observations recorded yet.
Historical observations retain their original currency. Prices are not converted or adjusted for regional taxes.
Memory & bandwidth
| Primary memory (comparison) | 147,456 GB coherent |
| Memory technology | Coherent HBM3 + LPDDR5X (256 superchips) |
| Memory per GPU | 96 GB |
| Total GPU memory | 24,576 GB |
| Coherent total | 147,456 GB |
| CPU system RAM | 122,880 GB |
| GPU / system bandwidth | — |
System memory, aggregate GPU memory and memory available to one accelerator are separate measurements. Do not treat total multi-GPU memory as a single unified pool.
Inference runtimes
No verified runtime compatibility records for this configuration yet.
Runtime support is not a model throughput or memory-fit benchmark.
144 TB GPU-addressable coherent memory combines 256×96 GB HBM3 and 256×480 GB CPU LPDDR5X. Not 144 TB discrete GPU HBM.