Functionally porting the linear algebra functionality of Ginkgo from CUDA* to SYCL*-enabled devices allows multiarchitecture, cross-vendor programming with the library. It also lets you use the latest Intel architectures, including Intel® Iris® Xe graphics and Intel® Data Center GPU Max Series.
With SYCL-ported Ginkgo on an Intel Data Center GPU Max Series 1550:
- Its sparse Matrix-Vector (SpMV) on average performs 2x better than the Intel® oneAPI Math Kernel Library compressed sparse row (CSR) matrix-vector implementation. The enhancement can even reach 100x for problems from the SuiteSparse Matrix Collection.
- Its batched iterative solvers on Intel Data Center GPU Max Series 1550 (one GPU, a single-socket system) on average run 1.7x and 1.3x better than on NVIDIA* A100 and H100 GPUs, respectively. Intel Data Center GPU Max 1550 2s (two GPUs, a double-socket system) outperforms NVIDIA A100 and H100 GPUs by an average factor of 3.1 and 2.4 respectively.