Nvidia boosts supercomputing speed with Tesla K40 graphics chip
- 18 November, 2013 14:03
Nvidia's Tesla K40 GPU accelerator
Nvidia wants to help create some of the world's fastest computers with its latest Tesla K40 graphics chip, which is its fastest supercomputing co-processor to date.
The Tesla K40 graphics processor delivers 1.43 teraflops of performance, which is roughly 20 percent faster than its predecessor Tesla K20X, which was announced a year ago. Nvidia's Tesla processors are widely used in supercomputers alongside CPUs to process complex scientific and math applications.
The K40 is "faster in every way" than the K20X, said Sumit Gupta, general manager of Tesla Accelerated Computing products at Nvidia.
The graphics processor is based on a modified version of Nvidia's Kepler architecture, which was also used in the K20X. The K20X chip was used in the U.S. Department of Energy's Titan supercomputer, which delivered 17.59 petaflops of performance and was the world's fastest supercomputer until June this year, when it was eclipsed by China's Tianhe-2, which has Intel's Xeon E5 chips and Xeon Phi accelerators, and delivers 33.86 petaflops of performance.
Tesla chips are also used in some workstations, but are not targeted at PCs. Cray, Hewlett-Packard, Dell, IBM, Silicon Graphics, Asustek and others will make products based on the K40, though Nvidia did not comment on the price of K40 or when it would be in supercomputers.
The K40 has 2880 processing cores, and 12GB of memory, which is double the memory of K20X. The base clock speed of all graphics cores is 745Mhz, which can be boosted to 810MHz or 875MHz, depending on computing needs. There is no ability to shut down idle cores as a power-saving measure, so the entire GPU is either lit up or not, Gupta said.
Nevertheless, the K40 is more power efficient than its predecessor, Gupta said.
There is an effort to boost supercomputing speeds as countries and companies race to reach the milestone of creating an exaflop computer by 2020. Chip companies are trying to reduce the power consumption of current chips and components, while boosting performance.