AMD GPUs for Scientific Computing: Reviewing the Software Ecosystem & Developer Experience

At present, NVIDIA dominates the consumer and scientific GPU markets with a market share above 80% for discrete GPUs. AMD GPUs look to be the most viable short-term competitor, but just how viable are they for scientific computing right now? In this talk, we will explore the state of the AMD GPU computing ecosystem and compare it against the NVIDIA ecosystem in terms of usability, library availability, support and performance. These aspects are explored through a publicly available suite of implementations of fundamental GPU algorithms e.g. parallel searches and sorts. These are implemented using both NVIDIAs CUDA and AMDs equivalent, HIP/ROCm. We will also discuss the developer experience of porting a more substantial code which is representative of a scientific computing application. Through this, feature availability is reviewed, looking at things such as atomic operations. In addition we review the state of AMDs compute library ecosystem, discussing the existence and/or effectiveness of equivalents to Thrust, CUB etc. Finally, we take a look at the viability of tools which offer automatic conversion of code from CUDA to HIP/ROCm. A recording of this session is available on YouTube: https://youtu.be/RM0B2nX5CZY