IBM's Power chips hit the big time at Google

The companies announced plans for the system, which they call Zaius, at IBM's OpenPower Summit in Silicon Valley on Wednesday. It's one of several new Power servers on show at the event.

They plan to submit the design to the Open Compute Project, meaning other companies will be able to use the design as well.

Google's data centers have long relied on servers based on Intel x86 processors, but the search giant has lately been exploring systems based on Power as well as ARM-based processors.

Two years ago, Google showed a Power server board it had developed for testing purposes, though it hadn't said much about those efforts since. It's now clear that Google is serious about using the IBM chip in its infrastructure.

"It won’t surprise anyone to hear that demand for compute at Google has been relentless, and it isn’t slowing down any time soon," the company said in a blog post. To meet the demand, Google's data centers need to be able to handle "ISA heterogeneity," it said, or the the ability to support multiple instruction set architectures.

That's a big change for Google, which historically has kept costs down partly by running the most homogeneous infrastructure it can.

The Power architecture is now "fully supported across our toolchain," said Maire Mahony, a hardware engineering manager at Google and director of the OpenPower Foundation. That means Google's developers can quickly deploy applications to Power systems.

She declined to say if the company is running production applications on Power today. But she said Google has ported "many" of its apps to the IBM chip.

Aaron Sullivan, a distinguished engineer at Rackspace, said cloud providers are attracted to Power for two reasons: One is that it's a good, high-performance CPU, and the other is that Moore's Law alone can no longer deliver sufficient gains from one generation of processor to the next.

To get the performance they need, companies like Rackspace and Google need the flexibility to rethink how their servers are designed, he said, including finding new ways to combine memory, I/O, and accelerator chips like GPUs.

That's easier with Power than with x86, he said, because IBM has opened the platform and removed licensing restrictions that otherwise make it hard for a community of customers and vendors to design new systems together.

The announcement continues a new level of openness at Google, which used to be quite secretive about the newest technologies it used in its data centers. It now apparently feels collaboration is in its best interests.

Last month Google joined the Open Compute Project and submitted the design for a 48-volt server rack that it co-developed with Facebook. The OCP is a place where end users and vendors collaborate on new infrastructure.

The Power server it's developing with Rackspace is designed to be compatible with the 48-volt rack.

The name Zaius comes from a character in the film "Planet of the Apes." The basic design is for a server with two Power9 processors and 32 DDR4 DIMM sockets.

Few other details were available Wednesday. Key to the design is that Power9 will support new high-speed interfaces, including IBM's CAPI and Nvidia's NVLink, that will make it easy to connect the CPU to accelerator chips such as GPUs.

IBM kicked off the OpenPower effort about three years ago. Aiming to breath new life into its struggling Power business, it opened up the platform to let third parties build servers and processors. The effort seems to be paying off.

Rackspace has already designed one Power-based server, called Barreleye, that it plans to put in production in the coming months, offering cloud services to its cloud customers.

Power is "the highest performance, most cost-effective option for workloads that run on Linux," according to Sullivan. For cloud customers, he said, Power servers running Linux "don't feel any different from any mainstream x86 system they've worked with in the past."

He dismissed the suggestion that companies like Google are showing interest in Power as a way to pressure Intel to reduce its pricing. "I so wish we could make that narrative go away, because it's so wrong," he said.

Intel isn't standing still, of course. The company has been selling custom Xeon chips to keep its cloud customers happy, and its acquisition of Altera will allow it to build accelerators into its own chips.

Along with the Zaius server, other companies also announced servers and components Wednesday that are part of the OpenPower effort.

Among them, IBM Itself is working with white-box server vendor Wistron to build a new HPC system that uses Nvidia's Tesla platform. The server will connect Power8 processors directly to Nvidia's Tesla P100 GPU via the NVLink interconnect.

Wistron's participation is significant, because cloud providers like Rackspace and Google like to buy systems from low-cost server makers in Taiwan and China.

Tyan, another white box maker, announced it has developed a 1U Power8 server. In addition, Mellanox, Xilinx and other companies announced new accelerators that work with IBM's CAPI interface.

More details about the new OpenPower products announced Wednesday are here and here.