OPEN LINK ↗
// 78d agoTUTORIAL
NVIDIA 595.45.04 patch unlocks RTX P2P
This Reddit guide shows LocalLLaMA users how to pair NVIDIA's 595.45.04 Linux driver with a patched fork of the open GPU kernel modules to re-enable peer-to-peer transfers on dual RTX 4090 and 5090 setups, including CachyOS-specific install steps. It is a niche but practical recipe for local LLM rigs that need faster multi-GPU communication on consumer hardware.
// ANALYSIS
This is exactly the kind of community workaround that keeps local AI moving faster than vendor support matrices. It is useful, but it also underlines how fragile consumer multi-GPU Linux setups still are for serious inference workloads.
- –The guide targets a real bottleneck for local LLM builders: inter-GPU transfer speed and topology limits on dual-card rigs
- –It relies on a forked kernel-module patch rather than an official NVIDIA feature, so maintainability and breakage risk are both high
- –The required IOMMU passthrough and related low-level tweaks make this better suited to dedicated inference boxes than everyday workstations
- –The post is most valuable for self-hosted AI operators running 4090 or 5090 pairs, especially on custom CachyOS kernels
// TAGS
nvidia-open-gpu-kernel-modulesgpuinferenceself-hostedopen-source
DISCOVERED
78d ago
2026-03-11
PUBLISHED
78d ago
2026-03-10
RELEVANCE
7/ 10
AUTHOR
Opteron67