Supermicro GPU SuperServer AS -8125GS-TNHR (HGX H200)
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 between GPUs; PCIe 5.0 x16 CPU-to-GPU
Notes
Supermicro's 8U AMD EPYC platform for the NVIDIA HGX H200 8-GPU baseboard, and the EPYC sibling of the Intel-based SYS-821GE-TNHR. Two SP5 sockets take AMD EPYC 9004 processors up to 128 cores and 256 threads at up to 400 W each, air cooled, with 24 DIMM slots supporting up to 6 TB of DDR5. GPUs connect to the CPUs over PCIe 5.0 x16 and to each other over NVLink with NVSwitch, and the chassis provides eight NICs for GPU-direct RDMA at a 1:1 ratio with the GPUs. The CSE-GP801TS chassis is 8U, weighs 102.1 kg gross, takes up to ten heavy-duty fans and six 3000 W Titanium supplies in 3+3 redundancy. Compute is identical to the H100 configuration because H100 and H200 are the same silicon at the same clocks; the difference is memory, 1.128 TB of HBM3e here against 0.64 TB. Supermicro publishes no FLOPS figure for this system; the compute shown is NVIDIA's for the HGX H200 baseboard it carries. Supermicro publishes no system power draw either, only the supply configuration, so total power is left blank.