- cross-posted to:
- [email protected]
- [email protected]
- cross-posted to:
- [email protected]
- [email protected]
There is hope at least. Currently ROCm is pretty terrible.
You must log in or register to comment.
If they could just help big distros package binaries that’d be great. I was unable to use their handy dandy installer to install on Debian 11 a couple of months ago, been meaning to try again now Debian 12 is available.
I believe Debian has official distro packages now. Arch, gentoo and NixOS certainly do, but they’re often a release or two behind. AMD only provides packages for the big corporate distros (Ubuntu, RHEL, SUSE), which I guess is fine. The odd one out is Fedora. There are official distro packages, but only for rocm-opencl, not for the whole stack. But ROCm is open source, so in the spirit of open source software, I believe distros should handle packaging duties. Only the distro maintainers know how they want to to compile, distribute and package the stack.
Also, this is what the average dev thinks of ROCm.
https://github.com/toverainc/willow-inference-server/issues/75
I like how they simultaneously complain that they’d have to use an “ancient” pytorch branch for intel while their example against rocm is that it apparently broke on a 4 year old kernel on the previous lts version of ubuntu.
Yea rocm has issues but many developers exaggerate a lot in order to justify not even trying.
My counter example is the easydiffusion ui. After users figured out that you only need to swap to the prebuilt rocm branch of pytorch and then it literally just works it took another couple of months (!) until upstream supported it. In the meantime I too used it with a literal 1 line change in the install script.
https://github.com/easydiffusion/easydiffusion/issues/115
The thing is, if nobody writes software for the alternative ecosystems, the alternative ecosystems won’t improve much.
If AMD wanted to get real serious real fast about rapidly improving drivers, could releasing Radeon driver source code under AGPLv3 help get mass development and fixes for drivers, while at the same time due to the license make the code untouchable for nVidia?
Both RadV and RadeonSI (both part of Mesa) are better than their Radeon driver these days. They should just abandon that unnecessary kludge and fully focus on the open-source drivers.
AMD currently has a market cap of $191B in an environment of very high interest rates and AI/GPU computing is already an existential issue for them.
The most sensible thing to do would be to raise $10B in cash by issuing stock and take advantage of the current AI/tech bubble and invest all of that money in the ROCm libraries to catch up with CUDA because otherwise they are currently at least a decade behind CUDA.
But noooo, Lisa Su will do nothing because shareholders will otherwise be angry at her for the dilution, and she’ll simply wait until the AI/tech bubble bursts and AMD loses more than $10B in market cap on its own.
