2012-03-27

Nvidia's Kepler to get compute cousin, says analyst

Nvidia's Kepler to get compute cousin, says analyst



MOUNTAIN VIEW, Calif.--Nvidia Corp. will likely announce a follow-up to its recently launched Kepler platform to better address upcoming compute products and the high performance computing (HPC) space later this year, according to an analyst.

David Kanter of Real World Technologies said the newly launched Kepler platform was specialized for high-end graphics, whereas more general purpose workloads would necessitate a differentiated version of the chip in order for Nvidia to remain competitive in the HPC space.

“It will be a derivative of Kepler, to re-use as much of the engineering effort as possible, but with several significant changes,” he said, hinting that a Kepler cousin could be announced as soon as May.

When it comes to different workload requirements, Kanter said it was clear that a graphics centric chip would not require much in the way of cross-system communication, whereas scientific computing did need it for algorithms to be efficient.

“For purely graphical use, the pixel that’s on the bottom left corner of your screen doesn’t care what the pixel in the middle of the screen is doing at all,” he explained, noting this wasn’t the case when trying to accelerate flow calculations or other more complex HPC data.

Thus, said Kanter, the hardware for each specific purpose -- gaming or scientific-- would have to be slightly different.

“Fermi had great resources for communicating between different parts of the application, but Kepler doesn’t have nearly as much capability to communicate between various levels of the system,” said Kanter, explaining that this would require a two-fold approach from Nvidia.

“I don’t believe Nvidia is abandoning HPC at all, it’s clear to me that the firm will continue on that path, because the area is really heating up, especially with AMD’s new offerings in the space and Intel on track to release Knight’s Corner at the end of this year. Nvidia can’t be trying to enter this market with one hand tied behind its back,” he said.

On the other hand, said Kanter, Nvidia is also acutely aware that it had not done quite as well as its rival AMD in the last round of the graphics battle, namely because it had spent a lot of resources in terms of power and area on things that were more useful for compute but not useful for graphics.

Hence, Fermi was much more attractive in the HPC space than in the niche high-end gaming segment.

“What Nvidia decided to do with Kepler was to scale it back, to really focus on graphics,” said Kanter, noting that the Kepler follow-up would likely be a bit more separate and distinct to fine tune it for general purpose.
Next: Compute specific GPU requirements

TAG:Nvidia Kepler GPU compute HPC

No comments:

Post a Comment