Original URL: http://www.theregister.co.uk/2010/04/22/google_the_server_chip_designer/

Forget the GPad - is Google building a server chip?

The needs of a 10 million node machine

By Cade Metz

Posted in HPC, 22nd April 2010 22:28 GMT

So, Google borged a mystery chip designer that was working on "some kind of server," and the web is convinced the Chocolate Factory is merely interested in using this all-star startup to build a GPad. How quickly the web forgets that Google is the world's fourth-largest server maker.

According to a New York Times source "familiar with the deal," Google acquired San José chip designer Agnilux not to build chips but to port its Chrome OS and Android operating systems to things like tablets and TV set-top boxes. And on one level, this makes sense. Agnilux was formed by ex-PA Semi employees once pulled into Apple to build SoCs for the iPod, the iPhone, and apparently the iPad, and it's no secret Google is exploring consumer devices well beyond its Nexus One phone.

But an earlier Times story indicated that Agnilux was brewing "some kind of server." The company apparently had a partnership with Cisco. And its roots can be traced back to server chips like the DEC Alpha and the AMD Opteron. It's been rumored for years that Google is interested in building server chips of its own, and if it hasn't already, you can bet that one day it will - with or without Agnilux engineers.

Google likes to say it's not a hardware company. When the ad broker launched the Nexus One, it went to unusual lengths to convince the world it played no part in the design of the physical device. But at the same time, it builds hardware on an epic scale for use across the Googlenet, a private infrastructure that handles more traffic than all but a pair of tier one ISPs.

It was recently estimated that Google runs 2 per cent of the world's servers, and it would seem that all of them are custom-built. Reports also indicate that Google builds its own routers, and it's no secret the company fashions its own data centers, piecing them together with hardware-packed shipping containers.

Last fall, Google released a brief video of a data center it built in 2005. The facility held 45 shipping containers, each housing 1,160 servers. Google is now operating about 35 data centers across the globe, and if you extrapolate, its total server count - server consolidation aside - is around 1,827,200. That figure is well above recent press estimates. And it may be low. After all, that data center was built in 2005.

According to a recent public presentation from the company, Google is intent on scaling its worldwide infrastructure to between one million and 10 million servers, encompassing 10 trillion directories and a quintillion bytes of storage. And this would span “100s to 1000s” of locations around the globe.

All these servers need chips. But the thing to remember about the ever-expanding Googlenet is that it's designed to process tasks that are broken into tiny little pieces. Google isn't interested in running the fastest processors on the planet. It's interested in running efficient chips that suit its pathological obsession with distributed computing.

When it released that data center video, Google also gave the world a peek at the battery-backed, two-socket server nodes it packs into at least some of its data centers. Based on a Gigabyte motherboard, each node included two disks, eight memory slots, and a 12-volt DC power supply. These nodes use both Intel and AMD chips, and it would seem that the company stops short of the bleeding x64 edge, choosing processors that provide the best performance per watt.

Plus, it wants chips that can run hot. To save costs - and, um, the planet - Google operates its data centers at temperatures above the norm. According to a former employee, at one point Google was buying chips from Intel guaranteed to operate at temperatures five degrees centigrade higher than their standard qualification.

What a Google needs

With an estimated 35 data centers and 1.8 million servers backing its search engine, advertising systems, and online applications, Google has power, cooling, and real estate problems that no other company in the world can appreciate. And with its custom designed GFS distributed file system and its MapReduce distributed number-crunching platform spreading tiny pieces of data across all those servers, its performance needs are outside the norm as well.

The PA Semi folks that founded Agnilux after leaving Apple have plenty of experience with chips and instruction sets of all sorts, and before Google acquired them, they could have been up to almost anything. All we really know the company is that it's based in San Jose and that its name is a mix of Sanskrit and Latin. Agni is the Sanskrit word for fire, lux the Latin word for light.

The assumption is that Agnilux was working on a variant of the ARM processor for use in servers, and given the thermal properties of ARM chips, this is a fair guess. But for all we know, Agnilux was building a variant of the Power chip - not the ARM. PA Semi was designing super-efficient Power clones when it was founded in 2003 by Dan Dobberpuhl, the lead designer on Digital Equipment's Alpha 64-bit chip, and Jim Keller, who worked on the Alpha and then the Opteron at AMD. Chip designers move from companies to startups and back again, and more times than not, they have great ideas that can't be commercialized. Think of the Transmeta chips that HP put into its first blade servers.

But with Google, building servers isn't a commercial undertaking - at least not directly. The Chocolate Factory can afford to indulge its whims, and it might as well build its own chips rather than waiting for Intel and VIA to tweak their Atom and Nano processors to handle Big Data applications on a small power budget. In this regard, Google would be acting more like a government-sponsored supercomputing lab, fashioning exotic hardware aimed at exotic problems.

And we might add that using low-power chips to take on high-power tasks is hardly a new idea. Witness the Atom servers from SuperMicro and SGI.

If this is indeed what Google is after, then the possibilities are intriguing indeed. It's easy to become a licensee of the ARM designs - it's quite possible that Agnilux already had the license - and parent company ARM Holdings would love nothing better than to have Google endorse its product over the x64 architecture. It would be the coup of the decade, and it will no doubt get other people looking at the ARM architecture for servers and desktops running Linux.

Because Google's workloads are based on Linux, presumably it's easy to port them over to ARM or Power or any other architecture that supports the open source OS. We don't know how tightly Google has compiled its applications down to the various x64 processors it uses. But, well, it just acquired a bunch of chip engineers and has plenty of software engineers to do a port to a new architecture if it thinks the benefits outweigh the hassle.

You can buy the GPad story if you like. But just for argument's sake, we point you to recent graphic from webhost Intac. It seeks to show - with big, bright colors - how many servers Google is running relative to the world's other tech giants (though it leaves out the likes of Yahoo!, Amazon, and Microsoft). Google is the one at the bottom. You can't miss it:

Google servers v the world

®