RTX 5090 MiniMax H3 performance degrades 8.7× at 2MP resolution
Suspicious-Walk-815 · reddit · 2026-08-30
A user running MiniMax H3 (ref2va, 20GB int8) via ComfyUI on an RTX 5090 (32GB) reports severe super-linear performance degradation. At 1MP resolution, per-step time is 30s (vs 21s reference), but at 2MP it spikes to 590s (vs 68s reference), an 8.7× slowdown. This suggests VRAM thrashing where the working set (model + activations) exceeds 32GB, causing CPU offloading every step. The setup includes PyTorch 2.12, CUDA 13.0, sageattention, and specific ComfyUI launch arguments. Adjustments like --reserve-vram 1 did not resolve the issue, and the process stalled at 1.5MP. The user seeks help diagnosing the configuration and valid workflows for RTX 5090.
Related event: RTX 5090 Suffers 8.7x Slowdown Running MiniMax H3(2 posts)→
More from coding & agent
- Grok Bot launches Templates to bundle skills, memory into shareable workflows — pswider · 2026-08-30
- Rote Playoffs hackathon opens: Agents save execution steps to cut costs — testingcatalog · 2026-08-30
- Computer Use Agent demo solves bot captcha automatically — cneuralnetwork · 2026-08-30
- AutoNodo runs 28 days processing 14B tokens, exploring massive context engineering — nodo48 · 2026-08-30
- Favorite AI slop: coding agents adding backward compat to an undeployed app — JFPuget · 2026-08-30
- Why Agents Editing Own Tools is Bounded Improvement, Not Recursion — dl_weekly · 2026-08-30