Disabling Smart Memory in ComfyUI Boosts Minimax H3 Generation Speed by 65-70% on RTX 3090
Life_is_important · reddit · 2026-08-10
A Reddit user shares a fix for using Sage Attention with Minimax H3 in ComfyUI on an RTX 3090: adding --disable-smart-memory resolved GPU memory management issues, cutting generation time by 65-70%.
- Issue: GPU VRAM filled to 23.5/24GB, power draw dropped from 230W to 110W, causing severe slowdowns.
- Solution: Disabling smart memory in ComfyUI reduced VRAM usage to 18-20GB and restored performance.
- Environment: Windows 11, RTX 3090 + 3060, Python 3.13.14, ComfyUI 0.31.0, comfy-aimdo 0.4.13, comfy-kitchen 0.2.28, PyTorch 2.13.0+cu130.
- Note: Manual update of comfy-kitchen was required; updating dependencies alone didn't help.
More from coding & agent
- Developer Automates 20 Business Tasks Overnight Using Multi-Agent Framework — leebase65 · 2026-08-10
- Human-AI Symbiosis: AI Handles Computation, Humans Retain Accountability — AryHHAry · 2026-08-10
- Build 3D Games in Seconds: Claude Code Integrates with Thrixel Platform — RanaHanocka · 2026-08-10
- Open-Source Skill: AI Analyzes Sports Data and Generates 3D Route Reports — op7418 · 2026-08-10
- Vibecoding with Claude Code: A Practical Guide for Non-Coders — Roger_M_Taylor · 2026-08-10
- How to Build a Shared AI Harness to Break Down Team Context Silos — VibeMarketer_ · 2026-08-10