Dual RTX 3090s run ComfyUI video gen 2.7x faster via new open-source Windows NCCL backend
unjusti · reddit · 2026-08-27
A developer benchmarked multi-GPU video generation in ComfyUI on dual RTX 3090s (80% power limit, 128GB DDR4) and built a working NCCL backend for Windows, developing much of it with Claude.
Key findings:
- For longer clips (160+ frames at 1MP), dual GPUs win big: 10s of 1280x768 renders at 25 s/it vs 67 on a single card, tested up to 13 seconds. Tested with Minimax-H3 ref2va INT8-pruned.
- For short clips and ref2v workloads with video references, a single 3090 with ComfyUI's built-in VRAM streaming is actually faster — avoid Ray + NCCL here; ComfyUI's dynamic loading is highly optimized.
- Sage Attention is 23% faster than Torch attention.
Two open-source repos:
- ncclwin: an NCCL backend for Windows usable with any torch-distributed workload; the DLL covers 20/30/40/50-series GPUs (only 30-series hardware-verified).
- ComfyUI-Raylight-Windows: a Windows patch for raylight whose install script fetches the prebuilt DLL, patches the node, and writes launch batch files with correct flags — every flag exists because something broke without it. Reports from dual-4090, 16GB, or 4-GPU rigs are welcomed.
More from Infra
- llama.cpp PR adds dspark support for Nanbeige4.2-3B — jacek2023 · 2026-08-27
- Oak CEO: Every AI agent needs distinct identity and auditability — TechNadu · 2026-08-27
- Structured generation with 0% throughput overhead in production — remilouf · 2026-08-27
- M5 Pro vs M6 for local AI: memory bandwidth debate resurfaces — jonejy · 2026-08-27
- Self-hosted MiniMax H3 renders for $0.03, compared to fal's faster H3 Max — tobowers · 2026-08-27
- Volcengine Unveils Embodied AI Data Solution with Seedance Video Generation — 火山引擎 · 2026-08-27