A Reddit developer running Qwen3 on four RTX 5060Ti consumer GPUs (16GB each) enabled a feature called PCI-E P2P (peer-to-peer direct access, letting multiple GPUs exchange data directly without going through the CPU). Prefill speed — the model's "reading speed" — jumped from roughly 1,649 tokens/sec to about 2,305 tokens/sec. That's a free 25% performance gain with zero hardware swaps.
What this is
P2P, short for Peer-to-Peer, lets multiple GPUs transfer data directly over the PCI-E bus, bypassing the CPU and system memory. It's a default capability on enterprise compute cards like the NVIDIA A100 and H100, but has long been locked out on consumer cards by the driver — a deliberate cut NVIDIA (the world's largest GPU company) uses to segment its pro and consumer markets.
This developer used a community-modified driver plus ReBAR (a low-level PCI-E spec that lets CPU and GPU exchange data more efficiently). Four 5060Tis give 64GB of total VRAM, running Qwen3-27B-FP8 (a 27-billion-parameter open-source LLM; FP8 is a low-precision compression format that lets the model run on less memory). Translation: roughly 20,000 RMB of consumer cards delivered a level of performance that previously required professional hardware.
Industry view
The community is circulating the numbers widely and framing this as a "poor man's H100" setup. Others caution that the 25% gain is measured in prefill; text generation speed (the model's "typing" speed) barely moves, and long-context workloads will see the benefit eaten by memory-bandwidth limits. The more sober read: single-GPU experience is unchanged, but the bar for multi-GPU coordination just dropped a real notch.
There's also pushback. The driver modification steps outside NVIDIA's official support, so the next driver update could break it. For 8x mode (PCI-E lanes halved, common on motherboards with multiple cards plugged in), the author admits they're "still working on it." This is a milestone result, not a steady-state conclusion.
Impact on regular people
For enterprise IT: if your company has strict data-compliance requirements and must deploy LLMs on-prem, you used to need professional cards or servers. A consumer-cluster setup using this method could cut procurement costs by more than half — worth evaluating.
For individual professionals: if you're learning AI and want to run models yourself instead of just calling APIs (cloud-hosted inference endpoints), an 8GB-VRAM laptop hits a hard ceiling. But a small desktop with 2–4 mid-range cards is becoming a credible new "good enough" starting point.
For the consumer market: gamers won't notice a thing. But this reinforces the idea that GPUs are no longer just gaming parts — and that NVIDIA's product-tiering strategy is being pried open, one tweak at a time.