dbanny OK, let me explain how AI/ML technology stack looks like. There are many layers involved:
GPU – Drivers – GPU Abstraction library for ML – ML framework – Application
Tonfotos is an application here, it does not deal with GPU directly. It only deals with ML framework. There are several popular frameworks, they are being developed by giants like Google (tensorflow), Facebook (PyTorch). etc. Most popular frameworks are free, that is why only such IT giants can affort to throw away hundreds of millions of dollars to develop it. We cannot, so we use one of available frameworks.
In turn ML frameworks do not deal with GPU or even GPU drivers either. This is too much of a hassle, all GPU’s are different with different capabilities, this is just creazy to keep updating your framework every time new GPU gets released. This is where abstraction libraries like CUDA come into play. They automatically detect what type of GPU you have, what it is capable of, and only use those functions that are availbale, while emulating all the rest.
NVidia does great job developing CUDA, and they are working with major framework developers to ensure they get best support possible and their frameworks work with CUDA flawlessly. They spend a lot on this. That is why CUDA/Nvidia is de-facto standard for ML/AI. There are no alternatives actually. ML frameworks just do not support anything except for CUDA. And therefore Tonfotos does not too, since it relies on one of the frameworks.
I am not really sure what exactly you mean by “AMD catches then with AI features”. I am sure they do something, but as far as ML/AI industry is concerned, they are still stuck at “GPU – Drivers” stage. They don’t offer anything comparable to CUDA for GPU abstraction. Until they do anything like that, and then convince all major ML frameworks vendors to support it officially, nothing will change.
Long story short, we are wrong people to complain to about AMD GPU’s support 🙂 This is definitely not our fault, and there is nothing we can do to change the situation.
