Google may be making its own custom server chips. It should

A thinly sourced report from Bloomberg Thursday speculated that Google is building its own custom server chips using the ARM architecture. This isn’t a new rumor, but something that has popped up every now and again since Google purchased ARM-based server chip design firm Agnilux back in 2010. I’ve tried to chase it down at varying intervals over the years.

After all, Google makes its own switches (although it did buy third-party chips for the boxes) and isn’t afraid to develop its own gear to optimize on performance and cost. When you make money delivering websites, it only makes sense that you are going to make the process as efficient as possible, and building servers designed to handle the relatively small workload of serving web pages would be an excellent use case for the lower-power ARM-based chips. Google has several other large-scale workloads where a custom processor could also make sense.

What’s changed in the last few years is the economics and power of ARM processors. As more companies are building out ARM-based systems designed for servers, the cost of building custom-systems on a chip has dropped. ARM is a modular architecture as opposed to one where aspects like networking are integrated into the chip. Thus building a custom chip using ARM and other elements doesn’t cost as much as it once did.

I even said this would come to pass in a story written this summer, when I quoted AMD’s Andrew Feldman discussing how large webscale companies would soon be building their own custom chips. From that story:

Andrew Feldman, GM and corporate VP at AMD, explained this idea in a series of conversations with me over the last few weeks in which he estimated that one could build an entirely custom chip using the ARM architecture in about 18 months for about $30 million. He compared this to the three or four-year time frame and $300 million to $400 million in development costs required to build an x86-based server chip….

… in the ARM world things are different. Because any number of players can license an ARM core, each one is looking for points of differentiation outside the core (some with architecture licenses are looking to tweak the core itself) and can make chips with better I/O or specific workload accelerators. An example here is Calxeda, which is using an ARM core in highly dense servers but has also built a custom interconnect to send information rapidly between its hundreds of ARM cores.

So when the mega data centers look at the opportunities presented by ARM, it’s not as simple as buying a piece of silicon from Marvell or Applied Micro, or a Calxeda box from HP. According to Feldman, web giants are looking at co-developing ARM-based chips that will take advantage of the greater levels of customization offered outside of the CPU so they can optimize for their own applications’ needs.

And as ARM has evolved its cores to meet the needs of the server market with 64-bit capabilities and an emerging server software ecosystem, webscale companies are evaluating the instruction set for everything from storage to data processing. Heck, AMD has even taken an ARM license.

So if Google is indeed considering custom ARM-based chips for certain workloads, that makes sense. For an investment in the few tens of millions it might be able to optimize workloads that could help it speed up its service or lower the cost of providing it. With a more modular and licensable IP core, if ARM can do the job, why not take a look at using it? When reached via email today, Feldman said he still believe that for Google it’s not a question of if, but of when, although he suspects we’d see chips coming out in the next three to five years. However, to make that possibility it would have to start designing today.

This might be a blow to Intel, which currently counts Google as its fifth largest customer according to Bloomberg, but my hunch is that Google would still use some x86-based chips in its hardware where that makes sense. This isn’t the religious war between ARM and Intel so much as its rationalization of the cost of computing.

When you get a large enough number of low-cost machines all doing the same thing — as you see in cloud computing or these large webscale players — placing all your bets on general-purpose computing is like depending on a Major League pitcher to also be a damn fine cleanup hitter.