At a recent Cambridge Game Creatures Meet-up in the UK, Andrew Richards, founder and CEO of Codeplay Software gave an overview of the multitude of different GPGPU programming standards being produced; the reason so many standards are useful; and what the challenges and opportunities are in defining GPGPU software platforms.
Andrew spends a lot of his time talking with both silicon vendors and software developers and makes some very interesting and at times amusing observations that many should find useful, I certainly did. 90 minutes well spent. Now you can share the experience thanks to YouTube and Cambridge Game Creators.
Paul Kier from Codeplay will be presenting at this year’s LEAP event, going into more detail on: The Impact of Low-Power GPUs (LPGPU) to Deliver Advances in Real-World Interactive Applications; Compiler Tools; GPU Power Modelling; and Hardware Design
Direct link: http://youtu.be/GjDdKR_C2io
There is so much in the presentation that its hard to pick out particular topics but a few things that I noted while listening were:
- Profiling – The need to teach it and for developers to use it effectively
- The reminder that the OpenCL spec states it’s not performance portable, but does it matter.
- That the GPU in a mobile processor can take up as much area as the CPUs
- The need for high-level abstractions
- The move to more SIMD GPUs, but interesting attempts to handle MIMD
- The different compute performance profiles of mobile GPUs (back to the old days of pc graphics)
Andrew Richards is CEO of Codeplay, chair of the OpenCL-HLM working group defining a C++ programming model for OpenCL and chair of the HSA System Runtime working group (defining a runtime system for the Heterogeneous Systems Architecture standard (“HSA”). Codeplay builds GPGPU compiler systems for semiconductor companies and have been working in the area of GPGPU compilers for over a decade.