I'm looking to upgrade my GPU for building and fine-tuning AI models, but I'm also into heavy gaming, though that's not my main focus. I'm debating between the RX 9070 XT and the RTX 5070 Ti. Both cards seem equally powerful, but the 5070 Ti comes at a significantly higher price. One big factor is CUDA, which is essential for AI training, and I've read that AMD's ROCm, its counterpart, is supported by PyTorch but may not be as efficient. Given that both GPUs have similar performance specs, I'm wondering if PyTorch will fully leverage the 9070 XT's capabilities, or if I should just go for the 5070 Ti.
2 Answers
I use a MacBook with MPS, and from what I've seen, CUDA is usually better and more refined for deep learning tasks. You might think about whether saving those extra bucks on the 5070 Ti is worth it for the performance you're after. For example, training a CNN that takes about an hour or so on my Mac took just 30 minutes on my friend's Alienware with CUDA. If you're fine with longer training times, the 9070 XT might still work for you.
In my experience, CUDA is definitely the way to go, especially if you're working with PyTorch. It's generally better optimized for most AI frameworks compared to ROCm. If the extra cost of the RTX 5070 Ti isn't a dealbreaker for you, it's worth investing in it for faster training times. But if you can live with slightly longer training sessions, the 9070 XT is still a solid choice.
Thanks for the advice! That really helps clarify things.
Good to know, thanks for sharing your experience!