AN ACCUMULATION OF ACRONYMS

Modern graphics cards are fitted with thousands of tiny processors, which is what enables them to compute so many shaders at once. These processors aren’t much use for some compute jobs but for others, where the work can be broken down into lots of units, they’re much more efficient than traditional CPUs.

Before Bitcoin ASICs became popular (and drove up the mining difficulty) GPUs used to be the most efficient way to mine them. Password cracking (or recovery) is another application where GPUs come into their own. John the Ripper and Hashcat both have options to use OpenCL to speed things up.

In addition to its open source driver stack, AMD has been working on a new open standard for GPU computing. Dubbed ROCm (Radeon Open Compute platform – apparently acronyms are allowed to do this now), which is aimed at high-performance computing (HPC), in particular machine learning (ML) applications. ROCm only supports newer AMD hardware (its Polaris graphics architecture and newer), with first-class status granted to its professional Instinct line of “GPU accelerators” (see https:// bit.ly/lxf292-amd-instinct).

ROCm backends already exist for popular machine learning applications, including TensorFlow and PyTorch. Blender can also make use of ROCm kernels in its Cycles rendering engine (see https://gpuopen.com/blendercycles-amd-gpu). This is good as the OpenCL backend was removed from Cycles last year.

Part of the ROCm initiative is AMD’s Heterogeneous-computing Intrastructure for Portability (HIP, this is allowed too) which hopes to aggregate CPU- and GPU-compute, as well as GPU-compute across different hardware. For example, it makes it possible to target either OpenCL or CUDA from a single codebase.