14 DAY TRIAL // After a long time when heavy computational tasks were exclusively handled by the Central Processing Unit, computing has begun to shift away from this traditional approach, with the latest Graphics Processing Units being capable of offloading tasks from the CPU. This method through which data is processed, in parallel, by both CPU and GPU has as the most widely adopted industry standard the OpenCL. While an accurate way of measuring the actual performance of CPU/GPU systems has not yet been developed, the OpenCL GPGPU benchmark suite aims to become the first application capable of such an analysis. The suite is a joint effort of AMD and SiSoftware and a part of SiSoftware Sandra 2010, the most recent incarnation of the award-winning utility invented in 1997.
The OpenCL benchmark suite is AMD's and SiSoftware's solution for those ISVs and OEMs that have been searching for a good method of measuring OpenCL-based configurations. The SiSoftware OpenCL GPGPU benchmark suite performs computationally heavy algorithms such as the Mandelbrot set. The chip maker is the only company currently providing a full OpenCL development platform for the entire system. Through this contribution, Advanced Micro Devices was able to reap special benefits, such as performance optimizations and solutions to certain issues through the use of the ATI Stream Software Development Kit (SDK).
To be more specific, AMD's ATI Radeon HD 5870 delivers, under some OpenCL benchmark tests, a 2.7 times faster performance than NVIDIA's CUDA-based GeForce GTX 295, which uses two GPUs, as opposed to the 5870, which only has one. For those interested in numbers, AMD's card scored 1,820 megapixels per second on the "native float shader" test, whereas NVIDIA's GTX 295 only managed 680.
Currently, the two partners, AMD and SiSoftware, are working on measuring the capabilities of AMD's entire platforms, including not only its GPUs, but also its x86 Central Processing Units. When complete, the OpenCL benchmarking suite will finally give the industry the instruments needed for accurate measurements and assessments of performance, allowing for better and more confident decisions to be made.
"AMD believes OpenCL is what the industry has been waiting for: an industry-standard, cross-platform development platform designed to allow developers to harness the immense computational power available in today’s GPUs and multi-core CPUs. We’ve been a staunch supporter of and contributor to OpenCL since its inception," Patricia Harrell, director of Stream Computing, AMD, said. "SiSoftware has made significant contributions to the OpenCL ecosystem with the release of its GPGPU benchmark suite with OpenCL support. This benchmark suite enables customers, partners and OpenCL developers to easily measure application performance on heterogeneous platforms, and provides the information required to help optimize this performance."
OpenCL supports four architectures, namely x86, x64/AMD64/EM64T, IA64/Itanium and ARM. The suite analyzes both computational and memory performance, supports parallel CPU+GPU execution (up to eight total devices), as well as multi-GPU installments (also up to eight). The suite works with both integrated and dedicated GPUs and on both AMD (AMD OpenCL 1.0) and NVIDIA units (nVidia OpenCL 1.0).
Information on the OpenCL GPGPU benchmarking tool may be found on the SiSoftware Sandra 2010 website.