Poster: On the Cost of a General GPU Framework: The Strange Case of CUDA 4.0 vs. CUDA 5.0

Abstract

The first release of CUDA was in 2007. Since then, we have experienced frequent new releases. CUDA reached its max performance with CUDA 4.0. Since its release, NVIDIA has started a re-design of the CUDA framework driven by software engineering prospective, i.e., the search for a general, multi-layer framework whose compiler back-end is unified with OpenCL. This can have significant impact on both maintenance costs and cross-platform portability. The software engineering community applauded this direction. However, our poster indicates that the new direction comes at a high cost in performance. Supported by the rigorous performance analysis of a MD code and its optimizations, we want to provocatively raise the question: Is the HPC community, that has been benefiting the most from CUDA and GPU programming, ready to face the challenge?

Related Papers

• → HIPCL(2020)21 cited
• → The portability of the BCPL compiler(1971)70 cited
• → On the Portability of CPU-Accelerated Applications via Automated Source-to-Source Translation(2019)12 cited
• → Observations on the portability of Ada I/O(1987)3 cited
• → Comparing Performance and Portability between CUDA and SYCL for Protein Database Search on NVIDIA, AMD, and Intel GPUs(2023)