Google is thought to be designing its own ARM chips. Peter Judge says, why not use IBM’s Power?
This is still a rumour, but it wouldn’t be a big surprise. We know that Google makes its own network hardware, and certainly already has its custom servers made. In 2010, Google bought Agnilux, a start-up founded by Apple designers which reportedly created a server chip, so it either has or can acquire the expertise. But is it ready to think inside the box -and design a server chip?
Google’s giant Intel spend
According to Bloomberg, Google is Intel’s fifth largest server chip customer, and provides about 4 percent of Intel’s revenue. Our picture shows the server racks in one aisle of one of Google’s data centres. Across the company it buys an awful lot of servers.
Take the money that Google spends on Intel processors, and there’s probably enough there to build a server chip. AMD struggles along with only a 4.4 percent market share in server CPUs, and it makes processors. Google has the benefit of a massive unified internal market that would use its chips, and could specialise significantly.
AMD’s Andrew Feldman told GigaOm that it takes $300 million to $400 million and four years to build a chip that uses Intel’s x86 instruction set, but things could be a lot cheaper on other designs.
British fabless silicon success ARM has a more customisable, modular design, and you can build a completely new chip with ARM cores in 18 months, for only $30 million, Feldman says.
That’s an investment that Google can afford to make.
What about Power?
But it doesn’t have to be with ARM. In August, IBM opened up its highly-regarded Power RISC architecture. Power is a success story, and powers IBM’s high-end systems, but those systems are a niche proposition, and the chips need bigger markets. IBM set up the OpenPower Consortium this year to license the Power architecture for servers and other users, in a pretty ARM-like way.
Since the announcement things have gone pretty quiet at OpenPower – but Google is a sponsor of the project and could be a marquee customer for OpenPower chip designs.
The drawback would be volume and experience. The ARM design is proven at the very high volumes required by smartphones, while OpenPower is still establishing this new business model.
Power has of course been used for high volume designs before – Apple adopted it for all its Macs a while back, before moving to Intel, and Microsoft’s Xbox 360 used a Power-based chip – before the Xbox One shifted to AMD.
It’s not a choice between the two, of course. ARM and Power may have different niches, and Google could have the size and skill to develop – or co-develop -two different processor chips.
Let’s wait and see.
A version of this column appeared on Green Data Center News.
What do you think of ARM? Try our quiz!