3 ms·
The fact that consumer and "pro"(?) GPUs don't use (mostly) the same software is not confidence inspiring. It means that AMD's already apparently limited capaci
by creato 2y ago
The fact that consumer and "pro"(?) GPUs don't use (mostly) the same software is not confidence inspiring. It means that AMD's already apparently limited capacity for software development is stretched thinner than it otherwise would be.
Also, if the consumer GPUs are hopelessly broken but the enterprise GPUs are fine, that greatly limits the number of people that can contribute to making the AMD AI software ecosystem better. How much of the utility of the NVIDIA software ecosystem comes from gaming GPU owners tinkering in their free time? Or grad students doing small scale research?
I think these kinds of things are a big part of why NVIDIA's software is so much better than AMD right now.
- wruza 2y agothat greatly limits the number of people that can contribute to making the AMD AI software ecosystem better I’d say it simply dials it down to zero. No one’s gonna buy an enterprise AMD card for playing with AI, so no one’s gonna contribute to that either. As a local AI enthusiast, this “but he used consumer card” complaint makes no sense to me.
- latchkey 2y ago> No one’s gonna buy an enterprise AMD card for playing with AI My hypothesis is that the buying mentality stems from the inability to rent. Hence, me opening up a rental business. Today, you can buy 7900's and they work with ROCm. As George pointed out, there are some low level issues with them, that AMD is working with him to resolve. That doesn't mean they absolutely don't work. https://rocm.docs.amd.com/projects/install-on-linux/en/latest/reference/system-requirements.html#supported-gpus https://rocm.docs.amd.com/projects/install-on-linux/en/lates...
- latchkey 2y agoAgreed that AMD needs to work on the developer flywheel. Again, not defending their software. One way to improve the flywheel and make the ecosystem better, is to make their hardware available for rent. Something that previously was not available outside of hyperscalers and HPC.