Nvidia Monday unveiled its Tesla K20 family of GPUs, targeted at the high-performance and supercomputing markets.
The new Tesla K20 chips are based on Nvidia's next-generation Kepler architecture, allowing them to reach processing speeds nearly three times those reached by GPUs based on the chip maker's prior-generation Fermi architecture. The Kepler-based design also makes the new Tesla K20 lineup more power-efficient than earlier GPU models, a feature that is crucial in high-performance computing environments, where intensive workloads must be handled quickly, without compromising efficiency.
Nvidia's Tesla K20 lineup consists of the K20 GPU, which touts 1.17 teraflops of peak double precision performance, and the higher-end K20X accelerator, which reaches an even greater 1.31 teraflops. The K20 was first teased by the chip maker earlier this year, when it debuted its new K10 GPUs for the defense and oil and gas industries.
The new K20X, according to Roy Kim, product manager in Nvidia's Tesla HPC business unit, can drive especially high levels of performance, making it an ideal fit for supercomputers in the life sciences and medical fields, where the use of compute- and graphics-intensive applications has become more and more common.
The K20X is being leveraged today in Titan, the world's fastest supercomputer, according to the U.S. Department of Energy, which hosts the machine at its Oak Ridge National Laboratory in Oak Ridge, Tenn.
"The GPU that is going into the Titan supercomputer is the K20X," Nvidia's Kim told CRN. "In terms of [the chip's] peak double precision performance [of] 1.31 teraflops, to put that in context, supercomputers back just even six years ago … entire clusters [of GPUs] performed at 1 teraflop of performance. And just six years later, we were able to condense that into just one GPU. So it's pretty stunning."
Titan is being used largely to track climate change and other weather patterns across the globe, while other supercomputers are leveraged to aid cancer and other medical research. According to Kim, GPUs such as the Tesla K20X are used to power these heavy-duty machines because traditional CPUs are simply unable to deliver the same levels of performance.
"In essence, CPUs are having a difficult time going faster because of what they call the power wall," Kim said. "People are turning to new technology such as accelerators and GPUs to solve their problems, and we caught the market at just the right time."
Systems based on Nvidia's new Tesla K20 GPUs will become generally available over the next two months, shipping from vendors including Hewlett-Packard, Fujitsu, Asus, Dell and IBM.
PUBLISHED NOV. 12, 2012