| Precision | Peak (PFLOPS) | Per GPU (TFLOPS) | TFLOPS/W |
|---|---|---|---|
| FP64 | 0.0 | 1.3 | — |
| FP32 | 0.6 | 75.0 | — |
| TF32 | 9.0 | 1125.0 | — |
| FP16 | 18.0 | 2250.0 | — |
| BF16 | 18.0 | 2250.0 | — |
| FP8 | 36.0 | 4500.0 | — |
| FP4 | 108.0 | 13500.0 | — |
| FP6 | 36.0 | 4500.0 | — |
All figures are dense. Vendors commonly headline the sparse number, which is twice the dense one.
NVLink 5.0 / NVSwitch 4.0 across the HGX B300 baseboard; 8x NVIDIA ConnectX-8 for InfiniBand XDR or Spectrum-X Ethernet scale-out
A 10U air-cooled single-node chassis carrying eight NVIDIA Blackwell Ultra SXM6 GPUs on the HGX B300 baseboard, which HPE names directly in the orderable part S6A37A. Two Intel Xeon 6 processors with P-cores drive it, either the 6776P or the 6767P, both 64-core at 350 W, with up to 32 DDR5 DIMMs and a minimum system memory of 2,048 GB running at 5,200 MT/s in the two-DIMM-per-channel configuration HPE recommends for performance. Scale-out is eight ConnectX-8 adapters supporting InfiniBand XDR or Spectrum-X Ethernet, with storage of eight E1.S drives and two M.2. Power comes from twelve fixed 3,200 W Titanium supplies in an N+N redundant arrangement, and HPE rates the machine at 14.3 kW continuous and 16.4 kW peak. Air cooling a Blackwell Ultra system of this density is unusual and it shows in the mechanicals: the chassis weighs 226 kg, ships pre-racked by default, and HPE requires a server lift on site before its engineers will service it. HPE describes it as delivering 50 percent higher FP4 performance than the prior Blackwell generation but publishes no FLOPS figure of its own, so the compute shown is NVIDIA's specification for the HGX B300 board rather than an HPE claim.