SYCLDB - cross-architecture acceleration of data analytics with SYCL
The growing popularity of AI and analytics in virtually all domains has led to an increase in the amount of data gathered. Workflows have become increasingly heterogeneous. While CPUs are not always efficient enough for the demands of modern AI and analytics workloads, GPUs, which offer greater parallelism, have largely remained closed-source. In response, two open standards have gained popularity: RISC-V, an open instruction set architecture, and SYCL, a cross-platform programming model.
Our work brings these two open standards together by developing SYCLDB, a SYCL-based library of key relational operations designed to run seamlessly across platforms. Leveraging the oneAPI Construction Kit (OCK), SYCLDB supports both RISC-V CPUs and accelerators, while also efficiently utilizing vectorization and multithreading on x64 CPUs and exploiting SIMT parallelism on NVIDIA GPUs. This flexibility does not degrade GPU performance, as SYCLDB achieves results comparable to its CUDA counterpart.
Speaker(s)
Ivan Donchev Kabadzhov - EURECOM, PhD student
Ivan is finishing his first year of a PhD in Data Science at EURECOM under the supervision of Professor Raja Appuswamy. His work focuses on heterogeneous, open-standard data analytics. In his day-to-day tasks, he develops cross-architecture database algorithms and investigates performance-portable implementation techniques on a variety of processing hardware. Previously, he worked on CPU performance profiling within the ROOT DataFrame at CERN.
