Dual 3090 Workflow: Prioritizing concurrency and utility for multi-modal pipelines
jbro1985 · reddit · 2026-08-21
The author shared a local AI workflow setup using dual RTX 3090s and a Ryzen 7, focusing on "utility" and "concurrency" rather than just raw speed.
Core Requirements:
- Needs to generate images, video clips, and scripts simultaneously.
- Requires concurrency to run multiple models (e.g., Code, Science, Hermes) side-by-side.
Specific Workflow:
- Image/Video: Uses Flux2 and Minimax H3 for overnight generation.
- Text/Review:
- Reduced max context for the 3.8 27B model to free up VRAM.
- Runs a Gemma MOE (offloaded) to assist with prose writing and proofreading Qwen outputs.
- Uses a smaller Vision Tower to automatically monitor the quality of Flux and H3 outputs during the night to prevent wasted runs.
Perspective: The author argues that optimization should be driven by the flexibility and overall utility of real-world use cases, not just single metrics like PPL or speed.
More from coding & agent
- AI coding shouldn't just patch bad designs; list standard solutions first — alexisgallagher · 2026-08-21
- o1 excels at fixing vision-grounded bugs in rendering pipelines — teortaxesTex · 2026-08-21
- Using Kimi K3 to write a fuzzer for the TypeScript compiler — DanielLockyer · 2026-08-21
- Secure local code inspection for ChatGPT Web via read-only MCP bridge — Lucaslogged · 2026-08-21
- We paid $550 testing x402 endpoints to build a verified directory for spending agents — SashSail · 2026-08-21
- MiniMax launches Design: an AI agent client for full-pipeline video creation — xiaohu · 2026-08-21