Intel's oneAPI / DPC++ / SYCL Will Run Atop NVIDIA GPUs With Open-Source Layer
Codeplay, which is already known for their several Vulkan / SYCL / SPIR-V initiatives, is working on this layer to run oneAPI / DPC++ / SYCL codes atop NVIDIA hardware while still leveraging NVIDIA's CUDA drivers. You could think of it akin as DXVK or VKD3D that map Direct3D 11 to Vulkan but this is about allowing Intel-focused code to run on NVIDIA's drivers. Or similarly, AMD's Radeon ROCm that allows some CUDA codes to be compiled for execution on AMD hardware.
Codeplay CEO Andrew Richards shared today that their support for the Intel APIs on NVIDIA GPUs with "CUDA underneath" will be "high performance" and work with existing NVIDIA products. Codeplay is expected to open-source this new project in early 2020.
And the second is that we are supporting SYCL, Data Parallel C++ and oneAPI on NVIDIA GPUs, using CUDA underneath. High performance, standard, on existing NVIDIA GPUs. We will open source in early 2020
— Andrew Richards (@codeandrew) November 18, 2019
It's possible that through Radeon ROCm and their CUDA portability initiative that might allow this Codeplay project to then run on AMD GPUs. But with Intel working to contribute their SYCL back-end to upstream LLVM and there already being the AMDGPU LLVM back-end and other components with both Intel and AMD building their graphics driver stacks from open-source, it's possible it won't be much of a headache in the end targeting AMD GPUs from Intel's APIs. It's NVIDIA that is the exception due to their closed-source CUDA ecosystem and not much to speak yet of open-source GPU compute on NVIDIA hardware.
Codeplay also announced at SC19 the "Acoran" standards-based platform focused on AI and high performance computing. This platform is based on OpenCL / Vulkan / SYCL / Intel oneAPI. More details on that initiative here.