Running ComfyUI and Local LLMs on a $360 AMD V620 GPU: Success and Benchmarks

Brave_Load7620 · reddit · 2026-07-31

A developer shared hands-on experience and benchmark data for building a local AI environment using an AMD Radeon Pro V620 workstation GPU (32GB VRAM).

Hardware & Cost: The total cost for the used card plus cooling modifications was about $360. Identified as gfx1030 architecture in Windows 11, it was successfully configured with ROCm, Triton, and Sage Attention, running stably alongside an RTX 9070 XT.

Performance: While not exceptionally fast, it is very consistent. In ComfyUI benchmarks, generating a 1080p image with Z-Image Turbo took 47s, and an SDXL anime-style image took 25s. Additionally, the Ace-step music generation model produced a 3-minute song in roughly 3 minutes. This setup is highly recommended for local AI users needing large VRAM for background LLMs while occasionally generating images or videos.

Related event: Running Local LLMs on a $360 AMD V620 GPU(2 posts)→

Original post →

More from Infra

Infra channel →