SiSoftware adds GPGPU testing to Sandra benchmark

Posted at: 3:11pm 10th September 2008 by Jeremy Laird

Sandra 2009 sorts the single-precision wimps from the double precision

The latest 2009 build of SiSoftware's familiar Sandra benchmark has been released with added GPGPU goodness, bringing it into competition with 3DMark Vantage, which tests a graphics cards ability to run gaming physics. From SiSoft's point of view, the new tests are a timely addition to the suite given that the general purpose compute potential of graphics chips has become such a hot topic over the past year or so. Along with Sandra's usual suite of CPU, storage, memory and platform tests, the 2009 version adds two GPGPU-specific benchmarks. Both NVIDIA's CUDA platform and AMD's Stream Processing are supported as well as up to eight GPUs running in parallel.

The first test is a raw computational benchmark which SiSoftware reckons will reveal how well your GPU handles general purpose instructions and data. For the record, it involves the generation of Mandlebrot Set fractals that are apparently quite the thing for simulating audio and video encoding, games and other applications where a single operation is applied to large data sets.

If that sounds like digital double speak, the key aspect of the test is that is allows CPU and GPU performance to be directly compared. One of the other intriguing features is support for both single (32-bit) and double (64-bit) floating-point precision. On existing graphics hardware, running double precision operations on GPUs has tended to incur a major performance hit. That hasn't been much of a problem to date given that games rarely make use of double precision. It's much more important for general purpose computing, however. Over time, we expect double-precision performance to be a key indicator of whether a GPU maker is really delivering graphics chip architected with GPGPU in mind, or just paying lip service to a fashionable subject.

Sandra's second test is a little more straight forward. The GPGPU bandwidth benchmark measures the memory bandwidth available to your GPU. It's derived from the stream memory benchmarks as used by Sandra's CPU memory tests.

In its typically efficient fashion, SiSoftware has crammed the entire benchmark suite into a compact 18MB download. It may be early days for GPGPU processing on the PC and as yet it's unclear whether this new synthetic test will provide worthwhile insight. But with Intel tooling up to release its own GPU-like co-processor, the Larrabee chip, late next year, and NVIDIA increasingly positioning its GPUs as the future of high performance general computing, anything that helps us get a handle on comparative GPGPU grunt is welcome.