Reproduce “Evaluating SYCL as a Unified Programming Model for Heterogeneous Systems” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Ami Marowka
This research introduces "singularity" as a three-dimensional metric (portability, productivity, performance) to evaluate SYCL's efficacy as a unified programming model for heterogeneous systems. The evaluation reveals significant inconsistencies, with performance varying by up to 66.9 times between memory models on CPUs and 42.7 times between kernel types on GPUs, demonstrating that SYCL's abstractions are not interchangeably efficient across diverse hardware.
Key insights
- SYCL's memory management abstractions (USM vs. buffer-accessor) are not performance-interchangeable; buffer-accessors consistently offer superior and more predictable performance across various workloads and hardware.
- Parallelism abstractions (NDRange vs. hierarchical kernels) exhibit strong performance dependencies on the chosen memory model and target device, with hierarchical kernels often showing suboptimal performance.
- While SYCL provides functional portability, its inconsistent performance across different abstractions and implementations, coupled with productivity hurdles from abstraction choices, prevents it from achieving true "singularity" for developers.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.