←back to thread

Ask HN: Why hasn’t AMD made a viable CUDA alternative?

183 points spacebanana7 | 3 comments | 01 Apr 25 14:37 UTC | HN request time: 0s | source

I appreciate developing ROCm into something competitive with CUDA would require a lot of work, both internally within AMD and with external contributions to the relevant open source libraries.

However the amount of resources at stake is incredible. The delta between NVIDIA's value and AMD's is bigger than the annual GDP of Spain. Even if they needed to hire a few thousand engineers at a few million in comp each, it'd still be a good investment.

Show context
postalrat ◴[01 Apr 25 14:41 UTC] No.43547347[source]▶
Because they dont want nvidia to be in control of their own development efforts.
replies(1): >>43547359 #
jsheard ◴[01 Apr 25 14:42 UTC] No.43547359[source]▶
The question was why don't they have anything as good as CUDA, not why don't they adopt CUDA itself.
replies(1): >>43547402 #
cmrdporcupine ◴[01 Apr 25 14:45 UTC] No.43547402[source]▶
Isn't the "goodness" of CUDA really down to its mass adoption -- and therefore its community and network effects -- not strictly its technical attributes?

If I recall, there are various "GPU programming" and "AI" efforts that have existed for AMD GPUs, but none of them have had the same success in large part because they're simply non-"standard?"

replies(2): >>43547524 #>>43549738 #
1. em500 ◴[01 Apr 25 14:54 UTC] No.43547524[source]▶
I thought OpenCL was supposed to be the "standard"? From the Wikipedia page, it's largely vendor neutral and not that much younger than CUDA (initial release Aug 2009 vs Feb 2007). Maybe some more knowledgeable people can comment why it seems to have been outcompeted by the proprietary CUDA?
replies(2): >>43548063 #>>43549504 #
2. the__alchemist ◴[01 Apr 25 15:36 UTC] No.43548063[source]▶
CUDA has a comparatively nicer user experience. If you would like to understand tacitly and have an nvidia GPU, try writing a simple program using both. (Something highly parallel, like nbody, for example)
3. fulladder ◴[01 Apr 25 17:36 UTC] No.43549504[source]▶
OpenCL was definitely supposed to be the standard, but as the sibling suggests, it's just much harder to use than CUDA. Also, it doesn't feel like OpenCL has much of a community around it, documentation doesn't seem great, and just the overall experience is very frustrating. I tried to implement something in OpenCL about 5 years ago, thinking it would be fairly trivial to port a simple compute shader from CUDA, and ended up giving up. Just too difficult.