PCIe 3.0 x1 isn't a bottleneck for tensor split: measured ~100MB/s vs 985MB/s limit
Introvertosaurus · reddit · 2026-10-27
A Redditor tested dual-GPU tensor splitting over PCIe 3.0 x1 in llama.cpp and found it far from a bottleneck. On 2× RTX 3060 12GB, switching from layer to tensor split boosted Gemma 31B Q5 from 14 to 20 t/s (+42%) despite one card on x1. On rented 2× RTX 5060 Ti running a 71B IQ3XS model, tensor split hit 20.7 t/s vs 11.7 (+77%), while nvidia-smi showed decode PCIe traffic of only 90–110 MB/s per direction—well under x1's 985 MB/s theoretical limit. Only model loading (4,000+ MB/s spikes) is slowed. Upgrading the motherboard or sacrificing an NVMe slot is likely unnecessary.
More from Infra
- Nvidia and Micron drove a third of S&P 500 profit growth last quarter — LexSokolin · 2026-10-27
- WSJ: Broadcom CEO sealed Anthropic's $63B chip debt deal with loan guarantees and higher rates — rohanpaul_ai · 2026-10-27
- Digital Sovereignty Is Now a Cloud Architecture Decision, and AI Pipelines Make It Harder — DavidLinthicum · 2026-10-27
- Houmo LQ50-24 M.2 accelerator claims 160 TOPS but real-world tests are scarce — lazy-kozak · 2026-10-27
- TRL's tree attention adds linear attention support with up to 3.46x speedup — SergioPaniego · 2026-10-27
- Tested: With Qwen3 27B and Dflash2, a Q4 Draft Model Gives the Biggest Speedup — milpster · 2026-10-27