AMD has divulged more information on their Open Compute Platform which deserves some attention. The ROCm officially supports OpenCL 1.2+ which consists of an OpenCL 1.2 compatible runtime and OpenCL 2.0 kernel language. Key features include Coarse Grain SVM, C11 Atomics, OpenCL 2.0 Images Support, Latency to compute optimisation and User Mode DNA.
On another note, the platform will leverage OpenUCX support for rich remote programming models. This next generation open source HPC communication framework is built on the foundation of MXM, UCCS and PAMI. Also, there’s broad industry support from leading companies including IBM, ARM, Mellanox, NVIDIA and of course, AMD. This is a rich platform for supporting MPI, OpenSHEMM and PGCAS.
AMD has achieved increased investment in key solutions for the Open Compute Platform. More specifically, the new initiative supports core math libraries like BLAS, FFT, RAND and third-party frameworks such as CHARM++ and HPX. Not only that, the ROCm supports various benchmarks and a host of other applications, some of them still in development.
Finally, the ROCm has added HIP to deep learning which modifies the widely used CAFFE machine-learning framework to make application porting less complex. Rather impressively, 99.6% of the code was unmodified and the porting time took less than a 1 week. Also, AMD’s solution supports all the major CAFFE features including multi-GPU, P2P, FFT, filters and more.
