HPE ProLiant Compute XD685 (HGX B200)
Overview
Performance Metrics
| Precision | Peak (PFLOPS) | Per GPU (TFLOPS) | TFLOPS/W |
|---|---|---|---|
| FP64 | 0.3 | 37.0 | — |
| FP32 | 0.6 | 75.0 | — |
| TF32 | 8.8 | 1100.0 | — |
| FP16 | 18.0 | 2250.0 | — |
| BF16 | 18.0 | 2250.0 | — |
| FP8 | 36.0 | 4500.0 | — |
| FP4 | 72.0 | 9000.0 | — |
| INT8 | 36.0 | 4500.0 | — |
All figures are dense. Vendors commonly headline the sparse number, which is twice the dense one.
Configuration
- Accelerator
- B200 SXM 180GB
- Count
- 8
- Memory per GPU
- 180 GB
- Bandwidth per GPU
- 7.7 TB/s
Power & Efficiency
- Aggregate Power
- —
- TDP per GPU
- 1000 W
- FP8 Efficiency
- —
- Aggregate Bandwidth
- 64.0 TB/s
System Details
Interconnect
NVLink 5.0 / NVSwitch 4.0 across the HGX B200 baseboard; 4x FHHL PCIe Gen5 x16 slots for fabric, NIC or DPU
Notes
The eight-way XD685 in its NVIDIA HGX B200 configuration, liquid cooled only, in the 5U DLC chassis S4R96A. HPE names B200 support twice in the QuickSpecs, in the supported-accelerator list and in the rule that the H200 and B200 liquid-cooled accelerators must be paired with this chassis, so the configuration is not in doubt. The accelerator kit part number is not recorded: the accelerator table lists two entries both described as HGX B300 with different part numbers, one of which is likely the B200 mislabelled, and guessing which would put an unverified number on a row buyers may quote. Two fifth-generation AMD EPYC processors, 24 DIMM slots and up to 6 TB of DDR5-6400. HPE publishes no system power draw, only twelve 3,000 W supplies, so nothing is summed; compute is NVIDIA's for the baseboard.