AMD RX 6800 diffusion slowdown: smaller models run 8x slower than bigger ones under ROCm
GuaranteePurple4468 · reddit · 2026-10-20
A Redditor documents a year-long performance anomaly running ComfyUI ROCm on an AMD RX 6800 (16GB): at 1280×720, the 6.6GB Illustrious XL hits 1.2it/s (75s for 50 steps) while the 3.89GB Anima takes 2.2s/it (109s) and 5GB MOSS 9.66s/it (490s) — smaller models up to 8x slower than larger ones. Upscaling widens the gap exponentially; smaller text encoders/VAEs change nothing. Video gen is worse: LTX runs 70s/it plus 15-20 minutes of tiled decoding (25+ min total), decoding spiking 40GB of system RAM. VRAM headroom stays healthy, so it's not OOM — likely model-specific ROCm kernel issues. Useful real-world data for AMD local-deployment users.
More from Infra
- PyTorch Conference 2026 opens in San Jose; Nebius and Nscale join as Platinum members — PyTorch · 2026-10-21
- Nebius joins PyTorch Foundation as Platinum member, deepening vLLM inference work — PyTorch · 2026-10-21
- Victor Taelin publicly begs to buy an 8-GPU Rubin machine — 'nobody is selling it' — fforres · 2026-10-21
- Nscale joins PyTorch Foundation as Platinum member after Anyscale acquisition move — PyTorch · 2026-10-21
- Decision Models launches on Antseed with an open-source auto model router — airesearch12 · 2026-10-21
- Gradium open-sources Phonon: 31M/40M edge TTS models with Rust inference stack for 10+ backends — l-m-z · 2026-10-21