Intel consumer motherboards can break PCIe P2P on multi-GPU AI rigs

Arli_AI · reddit · 2026-07-25

A Reddit post warns that Intel consumer platforms such as Z890 can be a bad choice for multi-GPU AI setups.

• The author says PCIe peer-to-peer (P2P) behaves incorrectly under the Arrow Lake PCIe root complex, which is a problem for inference and training workloads that depend on GPU-to-GPU traffic.
• They cite an NVIDIA open-kernel-module issue showing broken GPUDirect RDMA / BAR1 read behavior on Intel consumer hardware.
• In bandwidth tests with two RTX A6000 cards, enabling P2P without the patch led to unstable or halved throughput, and forcing patched P2P drivers caused vLLM tensor-parallel output corruption.
• The takeaway: for people building local multi-GPU machines, motherboard/platform choice can materially affect both performance and correctness.

Original post →

More from Infra

• LLM Serving Metrics Thread: Why TPOT and Uptime Make or Break User Experience — abhijithneil · 2026-09-11
• PlanetScale launches sharded Postgres: 768 servers acting as one, 1PB scale — dhruv2038 · 2026-09-11
• Can a 7900 XTX 24GB run Qwen locally? Reddit seeks ROCm tok/s benchmarks — thenomadexplorerlife · 2026-09-11
• RTK Terminal Compression Cuts Tokens but Leaves Your AI Coding Bill Unchanged — Bartaseth · 2026-09-11
• SF Compute founder: buying compute is 'an absolutely awful experience' right now — IgorCarron · 2026-09-11
• SmolVM open-sources persistent computer infrastructure for agents that outlive chat sessions — aniketmaurya · 2026-09-11

Infra channel →