QuixiAI Unlocks Consumer GPU P2P, Serves 262K-Context Qwen on 8 RTX 3090s
Eric Hartford of QuixiAI (author of Dolphin/Samantha) has released a fork of NVIDIA's open GPU kernel module (driver 610.57.04) that enables PCIe peer-to-peer (P2P) interconnect on GeForce consumer cards, verified on the RTX 3090 (GA102), 4090, and 5090.
Confirmed
- Eric Hartford located and fixed NVIDIA's software restriction on multi-GPU P2P direct connections for consumer cards and submitted a patch; however, he expects the change won't be merged upstream, suggesting it's an intentional artificial crippling by the vendor
- Using the fork's P2P interconnect, his open-source inference project SlimServe (based on ds4/vllm) served Qwen3.8-Flash-Next with a native 262,144-token context on 8x RTX 3090 (192 GB total VRAM, no FP8, no NVLink): 140 tok/s at concurrency 1 and 1200 tok/s at batch C32
- Running Qwen3.8-Flash-Next-NVFP4 (the nvidia quantized build) on 4x RTX 3090: 120 tok/s at concurrency 1 and 361.7 tok/s at concurrency 8
- Eric Hartford gave a detailed walkthrough of how the 8-GPU setup was built, emphasizing enterprise-grade long-context inference achieved with five-year-old consumer hardware
Not yet confirmed
- Whether the patch will be accepted into NVIDIA's official driver mainline remains unknown; the author himself is pessimistic
- Stability of the fork on larger deployments or cards other than the 3090/4090/5090 awaits further validation
Why it matters
- If NVIDIA really does limit multi-GPU performance on consumer cards via software, this fork offers budget-constrained researchers and small teams a low-cost path to long-context inference
- Eight used 3090s can host a 262K context at thousand-level tok/s throughput, dramatically lowering the barrier to large-context serving
2026-09-06 ~ 2026-09-07 · 6 related posts
Primary sources
- SlimServe runs Qwen 3.8 Flash Next on 8x RTX 3090s at up to 1200 tok/sec decode — QuixiAI · 2026-09-06
- [source] Eric Hartford's fork unlocks P2P for RTX 3090/4090/5090 on NVIDIA's open kernel driver — QuixiAI · 2026-09-06
- Dev Claims Fix for NVIDIA's Intentional P2P Nerf on Consumer GPUs — QuixiAI · 2026-09-06
- [source] Serving Qwen3.8-Flash-Next at 262K context on 8 RTX 3090s hits 1,200 tok/s — QuixiAI · 2026-09-06
- [source] Qwen3.8-Flash-Next-NVFP4 hits 120-361 tok/s on Quad 3090 via custom P2P driver — QuixiAI · 2026-09-07
- open-gpu-kernel-modules fork adds P2P for RTX 3090/4090/5090 consumer cards — QuixiAI · 2026-09-07