Open Source LLM

hipSYCL Seeing New Runtime For This SYCL Implementation For CPUs + ROCm/CUDA GPUs

The hipSYCL effort has been about supporting the Khronos SYCL single-source language built on C++ across any CPU with OpenMP as well as AMD Radeon GPUs via ROCm and NVIDIA GPUs via CUDA. The hipSYCL effort has a new "Lite" experimental runtime under development.

The hipSYCL effort is just one of several SYCL efforts in the ecosystem for running SYCL across CPUs, GPUs, and other accelerators. Here's a look via the hipSYCL project how their effort compares to others:


The hipSYCL effort is developed at Heidelberg University and is one of multiple exciting open-source SYCL efforts.

The newest hipSYCL work is a new "Lite" runtime version being developed. This new runtime is in good shape for CUDA, ROCm, and CPU back-ends with only a few caveats. This runtime is quite redeveloped compared to the existing runtime, supports having multiple runtime back-ends active at the same time to be used concurrently, a new memory management system, support for 3D data transfers with ROCm, and a whole list of new features.

More details on this new hipSYCL runtime via this GitHub pull request .