alt Hacker NewsI’m constantly baffled and amused on why AMD keeps majorly failing at this.
Either the management at AMD is not smart enough to understand that without the computing software side they will always be a distant number 2 to NVIDIA, or the management at AMD considers it hopeless to ever be able to create something as good as CUDA because they don’t have and can’t hire smart enough people to write the software.
Really, it’s just baffling why they continue on this path to irrelevance. Give it a few years and even Intel will get ahead of them on the GPU side.
Replies
I've never understood why they have such a fractured approach to software:hardware support. I remember reading and writing comments about this on hn nearly a decade ago now. It's a long time to keep making the same mistake.
They had the exact same kind of support issues back in the OpenCL days, where they didn't manage to provide cross platform, cross card support for same versions of the platform.
I have never been able to reconcile it with their turnaround and newfound competence on the CPU side.
> I’m constantly baffled and amused on why AMD keeps majorly failing at this.
i wonder if you've considered the possibility that there's some component/dimension of this that you're simply unaware of? that it's not as straightforward as whatever reductive mental model you have? is that even like within the universe of possibilities?
If I were Jensen, I would snap up all the GPU software experts I possibly could, and put them to work improving the CUDA ecosystem. I'd also spin up a big research group to further fuel the CUDA pipeline for hardware, software, and application areas.
Which is exactly what NVIDIA seems to be doing.
AMD's ROCm software group seems far behind, is probably understaffed, and probably is paid a fraction of what NVIDIA pays its CUDA software groups.
AMD also has to catch up with NVlink and Spectrum-X (and/or InfiniBand.)
AMD's main leverage point is its CPUs, and its raw GPU hardware isn't bad, but there is a long way to go in terms of GPU software ecosystem and interconnect.