GPU Performance Portability Using Standard C++ with SYCL
summary
Claims that standard C++ with SYCL runs on Nvidia AMD and Intel GPUs and delivers native-like speed are backed by concrete tooling and results: compilation via DPC++ compiler or ComputeCpp emits PTX or SPIR‑V, bundles device code into a fat binary, and can interoperate with CUDA kernels. Microbenchmarks show comparable throughput to CUDA/HIP, a large LAMMPS codebase (~800k lines) runs with only slight losses, and extensions add TF32 and tensor cores support for hardware‑specific acceleration.
AI-written summary · about
chapters (11)
- 0:00GPU portability with SYCL
- 1:49Speaker background and SYCL implementations overview
- 3:58SYCL in modern C++
- 6:49SYCL compilation & backends
- 10:19DPC++ toolchain flow
- 14:22Specialization vs abstraction
- 18:34Optimization trade-offs
- 21:00SYCL vs native APIs
- 23:27Productivity vs performance
- 25:13Key takeaways
- 25:57Q&A on SYCL interop
related talks
37mC++ for multi-accelerator, multi-vendor systems
01h 26mKhronos Sycl Language Framework for C++ Accelerators: Take Advantage of All the MIPS
56mA Smooth Introduction to SYCL for C++20 afficionados
01h 00mC++ on GPUs done right?
01h 22mCostless Software Abstractions for Parallel Architectures
01h 06mAsynchronous Computation in C++
01h 06mUtilize your CPU power
01h 03mThe Landscape of Parallelism