AMD breaks the teraFLOPS barrier

In the run-up to the International Supercomputing Conference (ISC), which opens Tuesday 17 June in Dresden, chip manufacturer AMD has announced its new FireStream 9250 application accelerator, which will apparently break the tera FLOPS barrier for single precision 32 bit operation.

According to AMD, the PCI Express card has a power consumption of less than 150 watts and does not therefore require an additional PCIe slot for a cooling system. According to the AMD figures, on performance per watt – an important measure for marketing purposes – the FireStream 9250 clocks up an efficiency of 8 GFLOPS per watt. This applies to single precision only, however. At double precision – 64 bit –, the successor to the FireStream 9170 achieves just a fifth of that performance, with a speed of around 200 GFLOPS.

By comparison, at a clock speed of 3 GHz, the 128 bit SSE units of one Core 2 Quad or Quad-CorGe Xeon achieve a total of 96 GFLOPS – 8 ops per clock cycle and CPU core – at single and 48 FLOPS – 2 fmul + 2 fadd each – at double precision. The nominal thermal design power (TDP) of two such Quad Core Xeons, which together achieve a theoretical maximum dual processor performance of 96 GFLOPS, totals 2 × 80 = 160 watts –. For its PowerXCell 8i (Cell eDP), a variation on its Cell chip, IBM cites a dual processor performance of 107 GFLOPS at a clock speed of 3.2 GHz.

The FireStream 9250 is presumed to be using the new RV770 GPU, which will also see action in the forthcoming Radeon 4800 series of graphics cards. AMD has not yet released technical details on clock speed and the number of shader ALUs. The only detail to have been revealed is that the memory will consist of 1 Gbyte of GDDR3 RAM.

The areas in which the General Purpose GPU (GPGPU) is likely to be used are HPC applications for financial analyses, seismic calculations and image processing. AMD will include its Stream SDK, which can be used to develop in-house applications for its co-processor. The chip manufacturer states that it is pursuing an open approach and, for example, offers the Brook + front end as open source. AMD has also announced that it is participating in the Khronos Group, which is currently developing the OpenCL open source GPGPU programming language. Apple is planning to integrate OpenCL into its forthcoming OS X 10.6 operating system. Microsoft is planning to release initial details of compute shaders in DirectX 11 at Nvidia's Nvision 08 conference.

The FireStream 9250 should be available for sale in the third quarter of this year priced at $999. Nvidia also sells an application accelerator in the form of its Tesla series and Intel is planning to enter the market in the next year with its multi-core Larrabee x86 graphics processor.

(trk)