I'm trying to decide between getting an Nvidia or AMD GPU to pair with my Ryzen 9950 X3D in a dual-boot setup (using Windows for gaming and office work, and Linux for heavy coding). I've heard that AMD generally performs better on Linux and that Sapphire is a solid brand. I'm also looking into whether I should go for consumer or workstation cards, especially considering the importance of Error Correction Codes (ECCs).
I primarily focus on JAX GPU acceleration for various tasks, not just neural network training, and both AMD's ROCm and Nvidia's CUDA are compatible with Linux. However, I'm cautious about spending over $1000 on a card only to discover that a cheaper alternative would have worked better.
Regarding workstation cards, I've been told they're often overpriced with less effective cooling. I'm also curious about what kind of errors ECCs actually correct. For instance, I experienced a situation where a CUDA-trained neural network started producing imaginary numbers, which shouldn't occur—could ECC have prevented that? Would ECCs be beneficial for my GPU-accelerated simulations and neural network training? The GPUs I'm considering are:
- Nvidia 5070ti
- AMD 9070XT
- Nvidia RTX Pro 4000 Blackwell?
1 Answer
What's your distro? I'm running Pop!_OS with a PNY RTX 5080 and it's been smooth so far. I rely on CUDA cores for my tasks, so that's why I went with Nvidia. They tend to work well in Linux environments, especially for gaming and creative work.
What are "CUDA cores" exactly? Are they special GPU cores that are optimized for running CUDA?
I mostly use Fedora, and I managed to get CUDA running on my 3060ti after some trial and error, so it's definitely doable with Nvidia on Linux.