Multi-node foundation-model training node
Lenovo positions the lower slots for eight NDR 400 Gb/s InfiniBand adapters with GPU Direct connections, the fabric this build uses to spread training across several servers. Two 64-core 8592+ processors keep four UPI links and a 5600 MHz memory bus, and all 32 slots take 128 GB modules for the 4 TB ceiling, enough host memory to stage large data sets.
- 2 × Xeon Platinum 8592+ — 64 cores @ 1.9 GHz, 350 W each
- 1 × NVIDIA HGX B200 8-GPU board — eight B200 SXM6 GPUs, 180 GB HBM3e each
- 32 × 128 GB TruDDR5 RDIMM — 4 TB, two per channel at 4400 MHz
- 16 × 7.68 TB PCIe 4.0 U.3 NVMe across both 8-bay backplanes — 122.88 TB raw
- Onboard NVMe ports, JBOD
- 8 × ConnectX-7 NDR400 OSFP 1-port adapters, one in each lower slot
- 8 × 3200 W Titanium CRPS Premium supplies, N+N
- Ubuntu 24.04 LTS
- Lenovo Premier Support, 5 years
Two DIMMs per channel hold memory at 4400 MHz rather than 5600. GPU-related onsite requests fall outside Lenovo’s service-level targets on this model, so onsite spares are worth pricing; single-person service needs a material lift, and Lenovo supports no more than five of these servers in a cabinet. Our configurator’s entry for this adapter carries a water-cooled (DWC) part name; the quote uses the ConnectX-7 NDR400 OSFP 1-port adapter Lenovo lists for this server (feature C51C).
Open this build in the configurator →