Dual V100 NVLink build asks which PCIe layout best serves local inference

In_der_Tat · reddit · 2026-07-21

• The author is planning a first local inference / light training setup around dual V100 SXM cards with NVLink, and asks which PCIe/backplane layout is better: a direct-through connection with two PCIe links, or a board that exposes the pair through a single PCIe connector.
• The post weighs the practical trade-offs: 64 GB VRAM, around 900 GB/s memory bandwidth, 300 GB/s GPU interconnect, lower upfront cost, but also no BF16 support and ecosystem obsolescence that the community partly mitigates.
• The broader build includes an Xeon E5-2699 v4, HP Z440 motherboard, DDR4 RDIMMs, a separate PSU for the GPU baseboard, custom cooling, and even 3D-printed housing. The author asks which configuration is most computationally and energetically efficient, and whether there are better alternatives.

Original post →

More from Infra

• NVIDIA pushes OpenUSD as the common layer for simulation and physical AI — MonaJalal_ · 2026-07-22
• SkyPilot exits stealth with $20M to unify fragmented GPU compute across five clouds — skypilot_org · 2026-07-22
• Production AI budgets include retries, routing, caching and observability—not just token prices — arx-go · 2026-07-22
• NVIDIA briefs analysts on Vera CPU and doubles down on monolithic agentic design — BenBajarin · 2026-07-22
• NVIDIA unveils Vera Rubin platform with claims of 10x better performance per watt — nvidia · 2026-07-22
• Why a 1GW Chinese AI data center may be plausible after all — teortaxesTex · 2026-07-22

Infra channel →