2-GPU R9700 vLLM rig hard power-offs: tensor-parallel all-reduce transient, not wattage, kills 1200W PSUs
brainchillzZ · reddit · 2026-10-19
A Reddit user debugged a dual Radeon AI PRO R9700 vLLM tensor-parallel rig that hard-powered-off 2s into every warmup, with zero kernel logs. The key finding: it's not total wattage. With cards uncapped at 300W and Radiance's custom one-shot peer-memory all-reduce (GPUs reading each other's VRAM over PCIe), it failed 11/11 times; falling back to stock RCCL or capping cards at 211W made it rock stable.\n\nThe controls are thorough: four different R9700s, a dedicated wall circuit, forced PCIe Gen4, single vs multi-rail OCK mode, and even a factory-replacement Dark Power 14 1200W PSU with new cables — which died at the exact same second. Meanwhile a 1000W Lian Li unit never failed. The puzzle: why multiple $300 Titanium-rated units can't handle this specific transient load pattern.
More from coding & agent
- Qwen Code v0.25.3 Ships Managed-Agent Fixes and Web-Shell Tweaks — qwen-code-review-bot · 2026-10-19
- Dad secretly installs GitHub Actions runners on his kids' computers for free CI — tristanbob · 2026-10-19
- Amazon Paper: QA Scores Miss Agent Memory Failures by 16.8 Points, New FEB-30 Benchmark Fixes It — rohanpaul_ai · 2026-10-19
- How One Founder Used Grok Bot to Submit Backlinks to 4000+ Sites in a Single Day — sujingshen · 2026-10-19
- Netflix shares how it built a knowledge base for production AI agents — AxSaucedo · 2026-10-19
- Reverse Engineer Anything crosses 300k GitHub stars as a movement — FinanceYF5 · 2026-10-19