HPE Cray XD670
Overview
Performance Metrics
| Precision | Peak (PFLOPS) | Per GPU (TFLOPS) | TFLOPS/W |
|---|---|---|---|
| FP64 | 0.3 | 34.0 | — |
| FP32 | 0.5 | 67.0 | — |
| TF32 | 4.0 | 494.5 | — |
| FP16 | 7.9 | 989.5 | — |
| BF16 | 7.9 | 989.5 | — |
| FP8 | 15.8 | 1979.0 | — |
| INT8 | 15.8 | 1979.0 | — |
All figures are dense. Vendors commonly headline the sparse number, which is twice the dense one.
Configuration
- Accelerator
- H200 SXM 141GB
- Count
- 8
- Memory per GPU
- 141 GB
- Bandwidth per GPU
- 4.8 TB/s
Power & Efficiency
- Aggregate Power
- —
- TDP per GPU
- 700 W
- FP8 Efficiency
- —
- Aggregate Bandwidth
- 38.4 TB/s
System Details
Interconnect
NVIDIA NVLink with NVSwitch across the eight SXM modules
Notes
A 5U server carrying eight NVIDIA H200 SXM GPUs at 700 W each with 141 GB of HBM3e apiece, driven by a single two-socket node of 5th generation Intel Xeon Scalable processors at up to 400 W, with 32 DDR5 DIMMs across eight channels per socket at up to 5,600 MHz. It is air cooled with a direct liquid cooling option. HPE has taken six first-place results with it in MLPerf Inference v5.1, including computer vision and LLM chat and text generation. Two things this row deliberately does not claim: HPE describes the GPUs only as "8x NVIDIA H200 SXM" and never names the HGX baseboard, though eight H200 SXM at 700 W is precisely the HGX H200 8-GPU configuration, and the compute figures shown are NVIDIA's for that baseboard rather than any HPE claim; and HPE publishes component TDPs but no system power figure, so total power is left blank rather than summing 8 x 700 W and 2 x 400 W, which would count silicon only and understate the machine. Part numbers are also not recorded, since none appears on HPE's page.