AMD and SiSoftware have teamed up to develop a benchmark testing suites for OpenCL. Released by SiSoftware, the OpenCL GPGPU benchmark suite is part of SiSoftware's Sandra 2010. OpenCL is an open standard for parallel programming of heterogeneous systems. SiSoftware's Sandra 2010 software provides remote analysis, benchmarking and diagnostic features for PCs, servers, mobile devices and networks.
The SiSoftware OpenCL Benchmarks focus on two major performance aspects -- computational performance and memory performance. Key features include:
- Four architectures natively supported (x86, x64/AMD64/EM64T, IA64/Itanium, ARM)
- AMD OpenCL 1.0 support
- nVidia OpenCL 1.0 support
- GPU + CPU parallel execution supported, up to eight devices in total.
- Different models of GPUs supported, including integrated GPU plus dedicated GPUs.
- Multi-GPUs supported, up to eight in parallel.
To test performance, the SiSoftware OpenCL GPGPU benchmark suite runs computationally intense algorithms like the Mandelbrot set.
Working with SiSoftware, AMD has optimized the performance of the OpenCL benchmarks for its GPU implementations, and for some problems AMD claims to have demonstrated performance advantages using AMD's ATI Stream SDK for OpenCL. When compared to NVIDIA's CUDA running on its GeForce GTX 295 featuring two GPUs, for instance, the ATI Radeon HD 5870 graphics card with one GPU delivers up to 2.7 times faster performance on certain benchmark tests. For the "native float shader" results, the ATI Radeon 5870 posted a score of 1820 megapixels per second, compared to the GTX 295 at 680 megapixels per second. AMD and SiSoftware are currently collaborating on measurement for AMD's entire platform, including x86 CPUs.