Researchers Boost GPGPU Performance by 20%

It has become the norm for many recent mainstream CPUs and APUs to have an integrated GPU on the same silicon. With technologies such as CUDA and OpenCL, it is possible to harness the computing of the GPU (both integrated and discrete GPUs) to perform certain, repetitive computing tasks. However, with current technologies, the CPU and GPU, even if on the same silicon, are rarely working together so are not as efficient as they might be. Researchers at North Carolina State University have developed a technique that allows GPUs and CPUs on the same silicon to work together, boosting CPU performance by an average of over 20%.

Although CPUs and GPUs are now being produced on the same silicon, they were designed to execute different types of task. GPUs were originally designed exclusively to display graphics, which meant they had to execute the same function on a set of data at very fast speeds. In contrast, a CPU can very easily perform more complex tasks (such as i/o tasks and sequences of non-repetitive calculations).

Dr Huiyang Zhou and Phd students Yi Yang and Ping Xiang from North Carolina State University have co-authored a paper with Mike Mantor from AMD, which details a novel approach to maximise GPGPU performance in fused CPU-GPU architectures. The paper, to be presented to the 18th International Symposium on High Performance Computer Architecture, New Orleans on Feb 27, describes a model where the CPU and GPU share L3 cache and off-chip memory (similar to current generation Intel Sandy Bridge CPUs and AMD APUs).

In the new architecture, the CPU is responsible for fetching data from main memory ready to be processed by the GPU. The CPU works ahead of the GPU, so the GPU can always find the data it needs stored in L3 cache, reducing program latency. Benchmarks run using this model indicate that the suggested pre-execution provides a performance boost averaging 21.4%.

AMD are committed to having an Open Standards Architecture for their HSA platform, so news and specs should be readily available. It should, however, be remembered that this is still just a simulated model, there is no silicon yet, so it is likely it would take a few years for a platform (including both silicon and software) to be available. The OSA does, however, mean that there is the potential for other compatible products to be available if this technology ever reaches the end user.