One x16 slot, eight GPUs: how PCIe-switch expansion boards use direct-connect vs cascade topologies
TheZachMueller · x · 2026-10-10
Zach Mueller breaks down a GPU expansion board with two internal PCIe switches that drives 8 GPUs from a single x16 slot. Direct-connect (two x16 to 4x x8 MCIO) routes through the CPU and suits CPU offload; cascade (single x16 into SW1, SW1 downlink feeding SW2) skips the CPU for inference, though cross-group traffic still tunnels through one x16. He plans a 9-GPU cascade build (8x RTX 6000 + 6000 ADA) and will benchmark TP8 vs TP4/PP2.
More from Infra
- AI value shifts from model layer to data center and apps as harness beats models, analyst argues — RihardJarc · 2026-10-10
- 5 open-source repos to build your own local AI apps: Ollama, llama.cpp, Jan and more — Shruti_0810 · 2026-10-10
- Samsung to adopt 4F² DRAM in 2028; CXMT to ship 4F² DDR5 RDIMM by year-end — zephyr_z9 · 2026-10-10
- Lumentum's AI optical components sold out through early 2029, CEO says — shashib · 2026-10-10
- SQLite core team builds Vec1: native ANN vector search is feature-complete — solyarisoftware · 2026-10-10
- SpaceX's four AI compute deals alone add up to $41B annualized revenue — XFreeze · 2026-10-10