OT: SYCL will replace CUDA

[Vattila]

Argonne, ORNL and Codeplay just announced that Codeplay has been awarded a contract for implementing SYCL support for AMD GPUs — as an open heterogeneous programming model for the upcoming exascale supercomputers.

Argonne, ORNL Award Codeplay Contract to Strengthen SYCL Support for AMD GPUs

LEMONT, Ill., OAK RIDGE, Tenn., and EDINBURGH, England, June 17, 2021 — Argonne National Laboratory (Argonne) in collaboration with Oak Ridge National Laboratory (ORNL), has awarded Codeplay Software a contract […]

www.hpcwire.com

I think this strongly signals that CUDA is not long for this world. What do you think? Since this is slightly off-topic here, head over to my poll in the Programming forum:

Speculation: SYCL will replace CUDA



Creating the same thread in different forums is against posting rules.
Just one thread is plenty.

Iron Woode
Super Moderator

 

[ThatBuzzkiller]

Are they going to implement a SPIR-V compute kernel compiler on AMD GPUs ?

If not then insinuating the decline of CUDA is premature when SYCL will have the same fundamental portability problems as OpenCL did ...

 

[Vattila]

Are they going to implement a SPIR-V compute kernel compiler on AMD GPUs ?

I am not sure what Codeplay plans to do for DPC++, but SYCL 2020 is fully independent of the backend. In fact, hipSYCL uses either the CUDA or ROCm toolchain (by compiling down to HIP, which is in turn compiled by the backend toolchain).

Head over to the Programming discussion thread.

 

[diediealldie]

I'd bet on CUDA, at least for a foreseeable time. One of the fundamental problem developers encounter is getting help from developer communities. SYCL? Not even close. When you're using CUDA with TF, Torch...etc then you're not alone. But if you're using some unknown backend compiled with no-one-knows hardware(FPGA??) then you're going to visit FPGA forum where only you and bots exist. It's not about money or time spent, it's more related to total project failure risk. This is how Intel ruled for 30 years. You're together with others. Easy to use.

Making people to believe that 'something will work and last forever' takes tremendous efforts. Capital and developers support. CUDA have been there for 14 years with only single company GPU(NVIDIA). Thanks to this investment, by the time CNN became so popular basically every graduates were already familiar with CUDA-based software. No one bothers using SYCL backend + Tensorflow today.

Just look at that diagram. ComputeCpp OpenCL + SPIR(-V). Who are responsible for that Intel CPU to Renesas R-Car compatibility. Even NVIDIA had trouble keeping everything on their hands. But now that much accelerators...yeah good luck.

If I have to, I'd bet on Intel's OneAPI. Intel knows hell about software and has everything on their hands so I think they can create one unified software environment which embraces FPGA, GPU(theirs and others) and CPU(Theirs!) together. But it's more like Intel's success, not SYCL's.

 

[tamz_msc]

CUDA got to be where it is today because for ten years and counting NVIDIA invested in supporting developers, making tie-ups with universities, getting top people in unis to create courses for CUDA programming, organizing workshops for students and conferences for researchers and so on.

Open source sounds cool but too often it is hampered by people with differing agendas trying to influence the direction of the project, resulting in a mess of standards and too many "ways of doing things" with the end result being an ecosystem that caters to a particular niche, with a handful of specialists who do their own thing but the public at large being unaware of the bigger picture.

 

[Vattila]

@diediealldie and @tamz_msc:

Head over to the Programming discussion thread for my replies.