AMD Announces The Radeon Open Compute Platform

AMD has just officially unveiled the Radeon Open Compute Platform (ROCm) which a new open-source foundation for Hyper Scale and HPC-Class computing. This initiative is based on a rich foundation of open source tools which tune deep learning frameworks and optimise parallel programming. As you can see, the HSA provides a host of features including user mode queues, architected queuing language, flat memory addressing and more.  The Rich Compiler Foundation and GCN headless Linux 64-bit driver are also of great interest which has been summarised below:

The project is a language-neutral solution to match developer needs as heterogeneous programming models evolve.

AMD’s Radeon Open Compute Platform has the ability to use a wider range of key coding languages which appeals to the entire engineering community. The slides mention OpenCL, CUDA, ISO C++ and Python.

The software has a comprehensive toolkit to simplify application development. For instance, the Heterogeneous Compute Compiler has a single source compiler for the GPU and CPU and support OpenMP 3.1 on the CPU. The Heterogeneous Compute Interface for Portability code can be executed on AMD or NVIDIA hardware and includes the right framework to simplify porting.

AMD’s latest creation is optimised to support dense GPU computing topologies.

Here we can see the timeline which resulted in ROCm 1.3 being released:

AMD has confirmed support for ROCm, Float16 and Integer 16 in GFX8 graphics cards. The native Float16 instruction support includes SUB, MUL, DIV, MADD, FMA arithmetic and COS, SIN, EXP, LOG, SQRT, FRACT, RCP in transcendentals. The slide also details the compiler hardware support and matrix multiply example.