Choosing a 32GB used GPU for local AI: V620, MI50, or V100
The purchase options are 32GB versions of the V620, MI50, and V100 s. The current computer has a 9070 XT , an MSI X670P Wifi motherboard, 32GB of DDR5 memory, and a Ryzen 7900X processor. The 9070 XT is currently used for ComfyUI and llama.cpp, including Gemma 4 26B A4B Q5.
The goal is to add a second card so the 9070 XT can stay free for gaming or other work. There may also be a later plan to combine the of both cards for larger . The needed setup is Windows 11, decent speed in ComfyUI for image models such as LTX 2.3, Z-Image Turbo, and Flux, plus acceptable and token speed in llama.cpp.
The core decision is which of the three cards fits this setup best and whether there are real benchmarks comparing them.
Key points
- The comparison is between 32GB versions of the V620, MI50, and V100.
- The current 9070 XT is being used for ComfyUI and llama.cpp but should stay available for gaming or other work.
- The target system is Windows 11 with an MSI X670P Wifi board, 32GB DDR5 memory, and a Ryzen 7900X.
- The main performance needs are speed, , and token speed.
- No benchmark numbers are included, so outside testing data is still needed.