Dual-GPU AI Workstation: R9700 for H3 + 3060 for Qwen Benchmark
Master-Client6682 · reddit · 2026-08-31
The author shares their experience building a dual-GPU local AI workstation with an AMD Radeon AI PRO R9700 (32GB) and NVIDIA RTX 3060 (12GB), focusing on MiniMax H3 video generation and Qwen LLM inference.
System Specs:
- CPU: AMD Ryzen 5
- RAM: 64GB DDR4 (Upgrading to 64GB was critical for H3; 32GB caused heavy swapping)
- GPU Setup: R9700 (ROCm 10) for ComfyUI/H3; RTX 3060 for Qwen/llama.cpp
H3 Performance (1MP Resolution):
Based on the last 50 production renders:
- Median Time: 6m 12s
- Mean Time: 6m 22s
- Slowest: 8m 12s
- Fastest: 6m 05s
- Massive improvement over the previous RTX 3060 (1 hour/video).
Qwen Inference (RTX 3060):
- Model: 27B Qwen (Quantized)
- Speed: 22 tok/s
Takeaways:
- ROCm 10 is viable but requires more tinkering than CUDA.
- System RAM is a bottleneck for H3 and must be scaled alongside VRAM.
More from Embodied
- Microduck: Hugging Face's $399 robot duck sells $2.6M in 24 hours — 创业邦 · 2026-08-31
- ModelBest Showcases Edge AI with SongGuoPi, Pushing Zero Marginal Inference Cost — 面壁智能 · 2026-08-31
- Toddler Cleanup Represents a Trillion Dollar Market for Robots — RachelVT42 · 2026-08-31
- Sim2real pipeline could revolutionize space robotics deployment — RachelVT42 · 2026-08-31
- Why avoiding potholes is harder than avoiding pedestrians for AI — aakashgupta · 2026-08-31
- VLANeXt codebase release reveals recipes for building strong VLA models — ccloy · 2026-08-31