467-run ComfyUI quant benchmark across 10 GPUs: int8 ConvRot most faithful, nvfp4 fastest but drifts
mybrianonacid · reddit · 2026-11-10
A Reddit user published a large-scale ComfyUI quantization benchmark: 5 storage formats (fp8 cast, fp8 scaled, int8 ConvRot, w4a8, nvfp4) vs bf16, across base and turbo versions of Anima, Z-Image, MOSS-Image, Krea 2 and SDXL/Illustrious — 15 models, 467 runs, timed on 10 GPUs (RTX PRO 6000, 5090, 5080, 5070 Ti, 5060 Ti, 3090, 3080 Ti, 3080, 3060 Ti) with identical seeds and prompts.
Key findings:
- int8 ConvRot is the most faithful format in every model tested, and on RTX 30 cards it's also fastest when it fits (2× bf16 on a 3090)
- nvfp4 is fastest on RTX 50 but drifts the most; w4a8 beat nvfp4 on accuracy in every model
- Never load a turbo LoRA onto an already-quantized model — most of its change rounds away; merge into bf16 first, then quantize
- 8GB cards run more than you'd think: Z-Image and MOSS fit entirely as int8 on a 3060 Ti at 2× bf16 speed; Krea 2 runs at 1.5 min/image
- Nothing failed to run: ComfyUI's dynamic VRAM streams from system RAM — costs speed, not crashes
- Quality tracks GPU generation, not the card: every RTX 50 matched the PRO 6000 within a few percent
- Odds and ends: PCIe x4 costs nothing once the model fits; monitors on the generating GPU cost a few percent; each 50W of power limit costs 4–6% speed for 3% energy saved; undervolting cost 5–10% speed for 20% less power
Missing: RTX 40-series — the author doesn't own one. The test kit is open-sourced (comfyui-quant-bench) and runs in Docker/WSL2 so others can contribute results. Full report with a "will my card run it?" lookup: ai-evolution.dev/quant-bench.
More from Infra
- Norway Suspends New Data Center Electricity Reservations Until as Late as August 2027 — Polymarket · 2026-11-10
- PagedAttention explained: managing KV cache like OS virtual memory — NielsRogge · 2026-11-10
- TSMC October revenue growth accelerates to +60% y/y on AI chip demand — firstadopter · 2026-11-10
- Archil claims self-hosted Spark on its S3-as-local-disk storage runs 2-5x faster than Databricks for under half the cost — massimosgrelli · 2026-11-10
- llama.cpp Adds --shard Flag for Multi-Location Model Loading and Separate PLE Files — pmttyji · 2026-11-10
- Four must-haves for agent sandbox providers, from hands-on setup — pramodk73 · 2026-11-10