by Anton Shilov
02/26/2008 | 12:14 PM
Nvidia Corp.’s chief technology officer David Kirk said at a news conference in San Francisco, California that in four years time graphics processors will be used in three out of five leading supercomputers on the planet. But before that Nvidia will have make friends with suppliers of top-performing supercomputers, many of which develop their own central processing units for high-performance computing apps.
<%BANNER[article]%>“By 2012, three of the top five supercomputers in the world will have graphics processors using parallel computing applications for computing,” Nvidia chief scientist David Kirk predicted, reports CRN web-site.
Both leading designers of graphics processing units (GPUs) – ATI Technologies (now graphics product group of Advanced Micro Devices) and Nvidai Corp. – have been talking about potential general purpose (GPGPU) computing power of their chips for many years. In the most recent 18 months both companies released their ATI FireStream and Nvidia Tesla GPU-based products for general-purpose computing applications, though, those devices have not become popular yet.
At present supercomputers are based either on specifically designed microprocessors or are powered by chips like AMD Opteron or Intel Xeon. Top supercomputers utilize hundreds of thousands of such chips that sport two or even more cores per each one. But modern graphics processors feature hundreds of unified stream processors, which means that theoretically GPUs can deliver ultimate performance. Potentially, this may allow supercomputer vendors to use lower amount of GPUs compared to the number of CPUs while delivering similar performance at lower hardware cost.
“This is truly the democratization of supercomputing. We ship a million parallel units a week,” Mr. Kirk is reported to have said.
Housed in Lawrence Livermore National Laboratory’s Terascale Simulation Facility, IBM’s BlueGene/L (BGL) supercomputer delivers 478.2 trillion floating operations per second (teraFLOPS) on LINPACK, the industry standard of measure for high-performance computing. Based on nearly 213 thousand of IBM PowerPC 440 chips, BGL has a peak speed of 596 teraFLOPS. Meanwhile, latest GPUs from ATI/AMD and Nvidia promise about 0.5 teraFLOPs performance. Potentially, about a thousand of graphics chips could substitute over 200 thousand of IBM’s CPUs, which is more than impressive.
One of the issues with GPGPU computing is that software has to be tailored for highly-parallelized execution on GPUs. Fortunately, due to multi-CPU nature of supercomputers, at least part of the job is done already. But there is still a lot of work to do with software.
Moreover, supercomputer and server companies like IBM or Sun Microsystems are not going to stand still and wait till Nvidia replaces their own microprocessors with its GPUs and will hardly embrace GeForce in cold blood. Therefore, both ATI and Nvidia will have to work hard before their graphics chips are utilized in high-performance computing applications and even then the success will hardly be completely inevitable.