Battle of Vega 20 vs Navi 21 32gb compute monsters!!
This page is dedicated to the AI performance differences in these two server grade graphics cards. I purchased two Mi50’s back in October 2025 and have made multiple YouTube videos regarding their usefulness in running local LLMs.
Today, June 2, 2026, I received my order from eBay of two Radeon Pro V620’s that were recommended from an eBay viewer. I will document their performance over the coming weeks. But will update this site as I can.
(rotate your device for better viewing)
The above charts display that running Vulkan backend, in Windows, the Pro V620 drastically outperforms the Mi50 in prompt processing, and does equal to / better than the Mi50 in most token generation situations as well.
Keep in mind, that the Vega gpus and prior work BETTER in HIP/ROCm scenarios when running inference.
Vulkan typically performs higher on newer cards… Navi and newer.
If you watch the video, you’ll see in the end that I had trouble getting Davinci Resolve Studio 21 to load. OpenGL / OpenCL had all sorts of issues when running either RX 7900 xtx or the RX 6800 along with the two Pro V620s.
Another issue that occurs, is that each Pro V620 also loads a “virtual desktop” that auto extends or duplicates your main screen. I right-click and enter Display settings, and deactivated screens 2 and 3… leaving only a single desktop via gpu1 (7900 or 6800).
If anyone has any tips on getting the 7900 xtx to play nice with the dual Pro V620s, please comment below!
Leave a Reply