A Taste of GPU Compute [video]

Re:A Taste of GPU Compute [video] by zozbot234 on Friday April 17, 2020 @11:08AM (#22899634)

Is Vulkan really becoming viable for compute? It seems like something that's obviously worth exploring given the overlap with more traditional graphics workloads, but I have seen some claims that it's still lagging behind OpenCL wrt. some features, and that it could be good for some things but is not obviously competitive across the board. It's possible that this impression is outdated though.

[ Parent | on HN ]

Re:A Taste of GPU Compute [video] by raphlinus on Friday April 17, 2020 @11:20AM (#22899766)

I think the answer to this is emerging, but I would say yes. Vulkan and SPIR-V are rapidly gaining features, while OpenCL is stagnating. Some of these features are very recent, like pointers and a memory model. Another area I'm expecting Vulkan to pull ahead is support for efficient float16 computation, which is important for ML workloads, and is increasingly well supported in hardware.

One thing to keep in mind though, programming directly in Vulkan is too low-level for most users, while OpenCL is a fairly friendly API. We need tools that can compile to SPIR-V and deploy on Vulkan, but high enough level you can express your problem rather than (for example) manually placing pipeline barriers between dispatches. The lack of such tools, and having to build some of my own, is one of the reasons for the slowdown in progress on my renderer project.

[ Parent | on HN ]

Re:A Taste of GPU Compute [video] by pjmlp on Friday April 17, 2020 @12:15PM (#22900271)

In CUDA I can just use ISO C++17 (C++20 around the corner) plus CUDA extensions, or any other language that targets PTX, with very nice graphical debugging from what is going on the GPGPU.

OpenCL is playing catch up with CUDA, but it requires at least version 2.0 for similar capabilities, while many vendors are still shipping 1.x drivers.

Vulkan, well it seems to be the same story as OpenCL regarding compute tooling, but lets see how much love it will get from NVidia.

I advise watching the GTC 2020 talks regarding CUDA and Vulkan compute.

[ Parent | on HN ]

Re:A Taste of GPU Compute [video] by maeln on Saturday April 18, 2020 @09:42AM (#22907651)

If you are looking for a single-source GPGPU framework, you should look for SyCL, not Vulkan or OpenCL.

[ Parent | on HN ]

Re:A Taste of GPU Compute [video] by Jasper_ on Friday April 17, 2020 @11:36AM (#22899930)

CUDA is unfortunately very, very far ahead of Vulkan. "Graphics Compute" is quite different from CUDA at the advanced levels if you've never used it.

[ Parent | on HN ]

Re:A Taste of GPU Compute [video] by jdc on Friday April 17, 2020 @03:13PM (#22902006)

Have you looked at SYCL yet?

[ Parent | on HN ]