Custom NVENC bridge splits FLUX 2 across two GPUs over LAN at 4.4s/image
Bypasses NVLink, uses Ethernet to combine a 5090 and 4090 for fast image generation.
A Reddit developer built a system to split large AI models like Flux 2 Dev and Klein 9B across two Nvidia GPUs over a network, using NVENC. Tested with a desktop at home and a laptop in a café via mobile tethering (Tailscale), generating a 1MP image in under 8 seconds (70% model on desktop, 30% on laptop). On a 1Gb Ethernet link between a 5090 desktop and 4090 laptop, Flux 2 Dev 1MP images took 14 seconds. Any Nvidia card with NVENC works. They plan to support LTX, Wan, and other large visual models. A separate codec for splitting 32B and 70B LLMs across two machines is ready and will be released this week.
- Achieves 4.4s per 1MP image by splitting FLUX 2 Dev across a 5090 desktop and 4090 laptop over 1Gb Ethernet.
- Works over WiFi 6 and mobile tethering—generates images in under 8 seconds using Tailscale VPN.
- Plans to release a similar codec for splitting 32B/70B LLMs across two machines; supports any Nvidia card with NVENC.
Why It Matters
Removes NVLink dependency, enabling consumers to run large models using multiple affordable GPUs over standard networks.