At the conference Supercomputing Conference 2016, AMD introduced Radeon Open Compute Platform or briefly ROCm. It will be the basis for modern heterogeneous HPC structures with HSA (Hyperscale) Accelerator for large-scale projects.
AMD has four base unit ROCm. The first is called Graphics Core Next Headless Linux 64-bit Driver. AMD mentioned earlier the development of drivers under Linux, here the company re-emphasizes the advantages of low delay, a smaller number of redundant computations and faster HSA on a discrete GPU. Access to hardware resources comparable to a modern graphics API Mantle and DirectX 12. The second important factor is the support and optimization of compilers under its own hardware resources. AMD has already shown the first solution in the form of Heterogeneous Compute Compiler as just a year ago. It represents the compiler with open source code under C++, allowing code to be compiled for the GPU. Supported and so-called HIP. It is a Converter that converts CUDA code in C++, which can then be run on the AMD GPU.
The third component is the integration of old and new structures of HSA, both software and hardware. Today, HSA is fully supported by modern GPUs and APUS, but still plays a significant role in the segment of GPU computing. But soon the situation should change, we can expect deep integration of the HSA. The fourth basic block tools and libraries open source, which is now available to developers.
The compiler plays a very important role in the new initiative because it would allow arbitrary code to run on AMD hardware resources as quickly as possible. Note that AMD already introduced the new LLVM compiler in the past year that can handle the code for calculations on the CPU. However, since the GPU gets a larger and more important share in the relevant calculations, the compiler GCN now provides a training code to a GPU. In particular, the compiler optimized architecture Graphics Core Next, the code obtains the corresponding optimization. AMD designed the interface for any code, including OpenCL, CUDA, ISO C++ or Python.
The first version of ROCm was introduced in April 2016, among other things, it may be noted support for GPU Fiji. Then in August, in ROCm added support for GPU Hawaii, and version 1.3 in November will get support for GPU Polaris. Also, the LLVM compiler is now released in final version. November should be supported by cards like Radeon Pro WX 4100, 5100 and 7100.
The idea of open source code aimed at attracting a wide number of manufacturers. First, the company plans to integrate its own architecture, Zen in the next year – of course, if you go out the first products. But discussing the prospects of AMD and other manufacturers, namely Cavium (ARM processors), IBM (processors Power8), and I/o – GenZ, CCIX and OpenCAPI Foundation. AMD is quite aware of the importance of deep learning segment, so the company’s efforts will be directed, including, and.