Looking beyond graphics processors, Nvidia is looking to push future Tegra chips into servers as the chip maker tries to break Intel's dominance in that market.
Nvidia is developing its first CPU for PCs and servers, code-named Project Denver, which is based on the ARM architecture and also aimed at mobile devices. The Denver core will go into future Tegra chips, and special improvements will be made to server chips, said Steve Scott, chief technology officer of Nvidia's Tesla product line of enterprise graphics chips.
"There are some things we are doing that are particularly nice for our purposes. It will likely go into the Tesla line at some point," Scott said.
Nvidia's current presence in servers is mostly related to its Tesla graphics processors, which are being used in the world's fastest supercomputers to perform complex scientific and math calculations. The Oak Ridge National Laboratory is building a supercomputer called Titan that will include Nvidia's Tesla processors and Advanced Micro Devices' 16-core Opteron CPUs to deliver a peak performance of up to 20 petaflops. The fastest supercomputer is Japan's K, which delivers a performance of 8 petaflops.
Scott did not share specific details on how Nvidia would tweak future Tegra chips for servers. However, the company has said that Project Denver chips will harness the parallel processing capabilities of Nvidia GPUs with ARM CPUs, which could boost server performance.
Most servers today run on Intel's Xeon and AMD's Opteron chips, but there is growing interest in low-power ARM processors as companies look to cut electricity bills. Analysts have said that while ARM processors may lack the performance and reliability to overtake traditional server chips for critical tasks, a large collection of lightweight ARM cores could process high volumes of Web-based transactions while drawing less power.
Running complex calculations by harnessing the parallel processing capabilities of CPUs and GPUs can speed up servers while reducing overall power consumption and computing overhead, Scott said. Nvidia is already building graphics cores in current Tegra processors.
"The ARM instruction set is more power efficient than x86. That's why there are people looking to build ARM-based servers. That's why we like ARM in phones, because you get more performance per watt, more performance per square millimeter," Scott said.
It makes sense for Nvidia to push its Tegra chips into the server market, which has higher margins than mobile devices, said Dean McCarron, principal analyst at Mercury Research.
"They have some interesting parallel processing technology that works out for them, and they have ARM, which makes sense for them to pair to go after that class of applications," McCarron said.
Nvidia's target market for server chips could be GPU-dependent systems delivering graphics or mathematical rendering in the cloud, McCarron said. ARM processors are not as proficient as GPUs in performing complex calculations, so Nvidia could end up making trade-offs on its ARM CPU design on power to bring in more performance.
"There are evolutionary pressures that drives you when you are going after servers compared to handhelds," McCarron said.
Companies like SeaMicro and Dell are building servers based on Intel's low-power Atom processors, but Nvidia's entry could fuel more interest in ARM servers. Nvidia's competitors will be Marvell, which last year announced a 1.6GHz quad-core ARM-based server chip, and Calxeda, which has built a server chip based on a quad-core ARM processor.
A big hurdle to entering the server market for ARM is software compatibility, as most of data-center code is written for x86 servers. A lot of IT implementations require corresponding server- and client-level compatibility, but x86 binary compatibility is less of a concern for Nvidia's future server chips delivering cloud services, Scott said.
"In the back room, in the cloud, binary compatibility doesn't matter nearly so much either," Scott said. "They are providing a service over the Web and they can switch to ARM, that is more power efficient."
The software stack is less of a worry on the server side than it is on the client side, where they could be issues around compatibility, McCarron said.
"As a user of a [cloud] service, the instruction set is meaningless. On the cloud side having to provide the service, that's where the investment comes in," McCarron said.
