Pros:<p><pre><code> Relatively low price - The 7900XT is more competitively priced compared to equivalent NVIDIA GPUs.
Good 4K gaming performance - For games at 4K resolution, the 7900XT can provide satisfactory performance.
George Hotz's tinybox project offers both 6x 7900XTX and 6x 4090, indicating that the 7900XT may have promising applications in certain computational fields.
Open-Source - ROCm (Radeon Open Compute) is open-source, while NVIDIA's official drivers are not. We should support open-source initiatives.
</code></pre>
Cons:<p><pre><code> Less mature software ecosystem - Getting research programs to run on the 7900XT may require substantial debugging and optimization work.
Higher power consumption - Compared to NVIDIA products of the same tier, the 7900XT generally has higher power consumption.
Driver stability issues - AMD GPU drivers have historically been less stable and reliable than NVIDIA's, potentially leading to driver crashes and other problems.
</code></pre>
Resources:<p><pre><code> https://github.com/geohot/7900xtx
Run machine learning on 7900XT and 7900XTX on PyTorch: https://gist.github.com/AlkindiX/9c54d1155ba72415f3b585e26c9df6b3
https://github.com/ROCm/ROCm</code></pre>
The consensus is that AMD cards work, but software support is pretty finicky. For my part, even if I wanted an AMD card, I'd either look for a great deal or wait until new nvidia cards just because they will likely push older card prices down. If you don't mind the headaches and papercuts of the software almost working.
Update: I think I'll just go with it. Maybe exchange/sell it later ;)<p>And thank you for your advice. I can't find a cheep and reliable marketplace for an old 3090 in China.<p>Our school provide some V100s on cluster, it may be tricky but able to run some program.
Not sure where you are located. If you're lucky enough to live in the US by a Microcenter computer store they often have refurbished NVIDIA cards for very reasonable prices. I picked up a 3090 for $699.