Three mini PCs over USB4 mesh run Qwen3-235B at ~22 tok/s with RDMA
AIFlow_ML · x · 2026-11-22
- AMD ran Qwen3-235B across three Ryzen AI Max mini PCs (128 GB each) linked via a 40 Gbps USB4 mesh, hitting 22 tok/s.
- The key enabler is RDMA over USB4 coming to Windows — no 200G NIC needed to split a large model across desktops.
- Numbers are vendor-reported for now; if they hold up, local distributed inference for big models gets dramatically cheaper.
More from Infra
- Reporter confirms data center noise is real: like a plane overhead that never lands — jdjohnson · 2026-11-22
- Small LLM judges hallucinated; tiny classifiers cut false cards to zero on phone — cyka_6lyat · 2026-11-22
- Zero data retention is trending: ZKPs for model computation go mainstream — yenkel · 2026-11-22
- OpenRouter essay exposes caching cheaters among inference providers, Novita flagged — AccBalanced · 2026-11-22
- Veteran engineer launching free FPGA course series to fix university gap — blaizedsouza · 2026-11-22
- Analyst questions SPCX's 10GW GPU order plan: prepayment financing clashes with NVIDIA's spot-market model — toptickcrypto · 2026-11-22