open-gpu-kernel-modules fork adds P2P for RTX 3090/4090/5090 consumer cards
QuixiAI · x · 2026-09-07
QuixiAI (Eric Hartford) released a fork of NVIDIA's open GPU kernel modules (driver 610.57.04) enabling PCIe peer-to-peer (P2P) transfers between GeForce consumer cards — RTX 3090 (GA102), RTX 4090 (AD102), and RTX 5090 (GB202). Inspired by George Hotz/tinygrad's P2P patch (later simplified by aikitoria) and ported to the 610 driver series. It also fixes two issues for unified-memory GPUs like the DGX Spark (GB10), verified on a Spark, including memory drain on last client close. Ships with install and compile scripts.
Related event: QuixiAI Unlocks Consumer GPU P2P, Serves 262K-Context Qwen on 8 RTX 3090s(6 posts)→
More from Infra
- M1 64GB Mac is no match for local Astra: 'this thing is crawling' — natesiggard · 2026-09-07
- KV Cache Engineering for LLM Serving: 12 Techniques Explained With Trade-offs — AccBalanced · 2026-09-07
- SlimServe open-sources low-cost LLM serving for consumer GPUs — QuixiAI · 2026-09-07
- AI buildout debt hits $570B, now rivaling the entire US muni bond market's annual supply — ivan_bezdomny · 2026-09-07
- 928-Star Wiki Details Running Qwen3.5-397B and Kimi-K2.5 on NVLink-Free PCIe RTX 6000 — TheZachMueller · 2026-09-07
- Prediction: compute is moving from rack-scale to datacenter-scale as bottlenecks shift outward — AccBalanced · 2026-09-07