←back to thread

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

183 points spacebanana7 | 8 comments | 01 Apr 25 14:37 UTC | HN request time: 0.001s | source | bottom

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.

1. Cieric ◴[01 Apr 25 14:55 UTC] No.43547535[source]▶
I can't contribute much to this discussion due to bias and NDAs, but I just wanted to mention, technically HIP is our CUDA competitor. ROCm is the foundation that HIP is being built on.
replies(3): >>43547554 #>>43550988 #>>43557565 #
2. johnnyjeans ◴[01 Apr 25 14:57 UTC] No.43547554[source]▶
I wonder what the purpose is behind creating a whole new API? Why not just focus on getting Vulkan compute on AMD GPUs to have the data throughput of CUDA?
replies(1): >>43548619 #
3. Const-me ◴[01 Apr 25 16:21 UTC] No.43548619[source]▶
I don’t know answer to your question, but I recalled something relevant. Some time ago, Microsoft had a tech which compiled almost normal looking C++ into Direct3D 11 compute shaders: https://learn.microsoft.com/en-us/cpp/parallel/amp/cpp-amp-o... The compute kernels are integrated into CPU-running C++ in the similar fashion CUDA does.

As you see, the technology deprecated in Visual Studio 2022. I don’t know why but I would guess people just didn’t care. Maybe because it only run on Windows.

replies(1): >>43551151 #
4. fransje26 ◴[01 Apr 25 20:28 UTC] No.43550988[source]▶
So if someone would like to, say, port a CUDA codebase to AMD, you would use HIP for a more or less 1-on-1 translation?

Any card you would recommend, when trying to replace the equivalent of a 3090/4090?

replies(1): >>43553577 #
5. Wumpnot ◴[01 Apr 25 20:46 UTC] No.43551151{3}[source]▶
Yes I found CPP AMP really interesting, but since it only ran on Windows..never used it for anything.
replies(1): >>43551763 #
6. Const-me ◴[01 Apr 25 21:58 UTC] No.43551763{4}[source]▶
It’s unfortunate they have deprecated it. We how have DXVK which implements D3D11, including compute shaders, for any platform which supports Vulkan. Making that (or a conceptually similar) thing work across platforms is no longer prohibitively expensive.

I believe that approach, i.e. the compute shaders, is the correct thing to do because modern videogames use them a lot, the runtime support is stable and performant now. No need for special HPC-only drivers or runtime components.

7. markstock ◴[02 Apr 25 04:05 UTC] No.43553577[source]▶
I can't recommend cards, but you are absolutely correct about porting CUDA to HIP: there was (is?) a hipify program in rocm that does most of the work.
8. stuaxo ◴[02 Apr 25 15:10 UTC] No.43557565[source]▶
OT: The thing where I have to choose between ROCm or AmdGPU drivers is annoying.

Mostly stick to AmdGPU as it seems to work for other stuff, I'd like to be able to run the HIP stuff on there without having to change drivers.