The problem has two sides: software and hardware. You can open source the software side all you want, it’s not gonna go very far when it has to fight against the hardware instead of working with it.
ROCm is open source, but it’s AMD. Their hardware has historically not been as powerful and therefore attractive to the target audience, so it’s been going slow.
The m chips are only good for inference and not for training. That is still unparalleled with CUDA. Pun intended.
I still don’t understand how an open source alternative with better hardware support hasn’t happened yet.
The problem has two sides: software and hardware. You can open source the software side all you want, it’s not gonna go very far when it has to fight against the hardware instead of working with it.
ROCm is open source, but it’s AMD. Their hardware has historically not been as powerful and therefore attractive to the target audience, so it’s been going slow.