The company aims at powering the top 5 supercomputers within a four-year timeframe

Feb 27, 2008 10:57 GMT  ·  By

Graphics expert Nvidia plans on taking over the supercomputer market in the next few years. The company has announced that it will power the top five supercomputers in the world. This move will give Nvidia's business a whole new dimension, and competing for the gaming market will become child's play. However, the company will have to struggle in order to get under the skin of the world suppliers of top-tier supercomputers.

"By 2012, three of the top five supercomputers in the world will have graphics processors using parallel computing applications for computing," claimed Nvidia chief scientist David Kirk.

The GPGPU (general purpose graphics processing unit) technology is currently mastered by both Advanced Micro Devices and Nvidia, the most important players on the high-end graphics market. In fact, both manufacturers released GPGPU-enabled products: the ATI FireStream and Nvidia Tesla, but the devices have not become too popular as of the moment of writing.

The current supercomputer technology is built around the regular x86 processors in considerable amounts. These chips feature two to four CPU cores each, which allows a supercomputer to boost its computing power.

However, the graphics chips come with hundreds of unified stream processors, which means that they can deliver improved performance over the dual-core models of AMD Opteron or Intel Xeon chips. This would ultimately account for less GPUs to act as CPUs inside a supercomputing cluster, thus decreasing the necessary physical space and the electricity bill, not to mention the costs for processing hardware.

"This is truly the democratization of supercomputing. We ship a million parallel units a week," Kirk said.

The practical approach is as follows: world's most powerful supercomputer, the IBM BlueGene/L can perform 478.2 TeraFLOPS per second on its 213,000 PowerPC 440 processors. In contrast, the latest graphics processors from AMD and Nvidia - the RV680 and the GeForce 9800 GPUs can deliver 0.5 teraFLOPs performance. It does not take rocket science to see that about 1,000 graphics chips will substitute for more than 200,000 x86 processors.

The catch is that GPGPUs won't "understand" the x86 software, which will have to be ported for the new architectures, and this requires a lot of efforts from the operating system and applications makers.