Asked if Intel customizes microprocessors for its biggest customers, Diane Bryant said: "Yes."
Bryant heads the Intel group that builds server chips and other hardware bound for large data centers. Last week, during a dinner with reporters in downtown San Francisco, she was explaining just how much the server business has changed in recent years. In 2008, three server giants -- HP, Dell, and IBM -- accounted for 75 percent of the revenues Intel pulled in from the sale of server chips. But today, Bryant said, that 75 percent is spread across eight server makers, and one of them is Google, a company that only makes servers for itself.
Then a reporter asked if Intel customizes parts for its largest customers. "Yes," Bryant said. "We want to give them a way of differentiating their machines."
Bryant didn't say much more, but those few words shine a light on another part of the big-time chip business that's rarely discussed. There are cases where a large chipmaker such as Intel and AMD will provide certain customers with chips that others may not have access to. Sometimes, this merely means that when the chipmaker cranks out a big batch of processors, one customer gets the chips that happen to have the best speed or power ratings. But in other cases, the chipmaker will actually modify processors at the request of a particular customer.
This practice may show how determined the Dells and the HPs are to offer machines that stand out in what has become a commodity market. But it may also show how far Google and other web giants will go as they work to customize the servers that underpin their online services, pushing to reduce power and cost in the data center.
Intel declined to provide additional information about its efforts in this area, and Google declined to comment as well. But rumors have long held that Google pushes for more than just ordinary chips from Intel. Intel rival AMD says it has customized chips in certain cases. And Facebook -- which also designs its own servers -- acknowledges that it requests specific silicon from the big chip makers.
According to John Williams, vice president of server marketing and business development at AMD, there are situations where his company has turned off certain parts of a processor at the request of customers -- or even added "instructions," the fundamental operations that define how a processor works.
Asked if Facebook makes such requests, Facebook spokesman Michael Kirkland said: "We do work with Intel and other vendors in these ways." But he stopped short of providing details, saying the company had not yet asked chip makers for approval to discuss the matter with the press.
As the big web companies step up their efforts to remake the hardware driving their massive data centers, chip customization could become a key battleground in the ongoing microprocessor wars. In recent months, Intel has acquired several technologies that would allow it to heavily customize server chips, but its rivals question whether it has the freedom to do so.
Intel operates massive chip fabrication plants, or fabs, that have traditionally been geared towards the production of millions upon millions of largely identical chips. According to Andrew Feldman -- co-founder of a new-age server maker called SeaMicro, which was recently acquired by Intel rival AMD -- there are other companies better suited to the creation of customized server chips for the big web players. Yes, one is AMD, which no longer runs its own fabs. The other is ARM.
ARM is the company behind so many of the chips that drive today's smartphones, but it's slowly moving into server chips as well. Though its new server chip designs are still a long way from live data centers, they've piqued the interest of many tech outfits because they consume relatively little power, a major concern for the big web players. But there may be an added attraction: ARM doesn't build its own chips. It licenses chip designs to others so that they can then be, well, customized.
There's Custom. And Then There's Custom
Andrew Feldman says it's only natural for big-time computer makers to push for chips that suit their particular needs. "Customers are always asking for features and performance characteristics from vendors. It is no different in the processor business," he says. "As a system vendor, you have deep insight into the end customer's requirements. You try to convince your CPU vendor to listen to you and put the features you think the customers will want into the chip."
But there's a wide spectrum of ways this can play out. "Custom can mean a lot of things," says AMD's John Williams. Sometimes, he says, when a buyer receives "customized" chips, this merely means they're getting the best silicon that comes off the line. Though incredibly precise, the modern chip manufacturing process ends up producing processors with a relatively wide range of characteristics, and based on these characteristics, the chips are sorted into separate "bins." Certain customers, Williams says, will get access to the bins with the fastest chips -- or the chips that consume the lowest power.
>'Customers are always asking for features and performance characteristics from vendors. It is no different in the processor business.'
Andrew Feldman
Separately, customers will ask for specific changes to a chip. A computer maker may convince an Intel or an AMD to build a new version of a chip that's then offered up to the market as a whole, but it could also push for chip customizations that aren't available to the rest of the world.
Andrew Feldman's SeaMicro began life as a company dedicated to building a new breed of low-power server using Intel's Atom chip, a processor originally designed for smartphones and other mobile devices, and according to the company, Intel built a new version of the Atom -- the N570 -- with "significant and specific input" from SeaMicro engineers.
This is an example of a customized chip that's available to the market at large. And since releasing the N570, Intel has built yet another version of the Atom known as the Centerton, saying it's specifically built for servers. HP has now said it will offer systems using the chip. "Intel often works with OEMs to define...custom chips, based on trends we're seeing from our mutual customer base," says Intel spokesman Mark Miller, referring to "original equipment manufacturers" such as Dell and HP.
But sometimes, chip designers will go further, customizing chips just for one particular vendor. According to AMD's John Williams, the company has turned off certain parts of a chip at the request of customers looking to save power, and in other cases, it has actually enhanced a chip. "There are certainly instances where a customer will say: 'Can you integrate this particular piece of our intellectual property that we have to differentiate our system and we want to be the only one who has access to it,'" he says.
Williams adds, however, that this sort of customization is rare. When AMD goes so far as to add instructions, it will typically make these changes available to all customers. For AMD, that just makes more economic sense.
Feldman says much the same thing, though he believes chip customization is even more difficult for Intel because -- unlike AMD -- the company runs its own fabs. And this will become more of an issue as customers request additional customization. A system maker could convince an Intel to add an instruction or two to an existing processor, Feldman says, but arranging for a completely custom chip is a different matter. The fabs run by a company like Intel, he says, are geared towards the production of large numbers of identical parts, not custom-built processors.
"Fabs are like printing presses," he says. "They are at their best when they are knocking out a lot of identical parts. They are at their worst when they are doing many different jobs. Because changing jobs takes time, and the press is only making money when it is running, not when it is setting up new jobs."
A New ARM for Google?
As Feldman points out, ARM is a very different animal. ARM's entire businesses, he says, is geared towards "custom" chip designs. The company licenses its core designs to outside outfits who then massage these designs to suit their particular needs.
Feldman believes that chip customization will become increasingly important in the server market. This, he says, could make the ARM architecture particularly attractive to the big web players.
"ARM has more than 100 of these licensees, each of whom is doing a 'custom' part. These chips are almost always customized for a specific application -- one particular phone or some such," Feldman says. "This chip could be customized for a very large data center customer just as easily -- or a very large data center company could become an ARM licensee and make their own part."
>'Our business model allows this to happen. But there would have to be quite a lot of ground covered.'
Ian Ferguson
At the moment, that seems a bit of a stretch. ARM chips still lack some key ingredients needed to run servers, and even ARM says we're long way from the Googles and the Facebooks building their own processors. "Our business model allows this to happen," says Ian Ferguson, ARM’s director of server systems and ecosystem. "But there would have to be quite a lot of ground covered from where those [web companies] are today to a situation that would make it a reality."
As it stands, he says, ARM has not licensed technology to the big web players.
Ferguson is also careful to say that ARM machines still need a lot of work before they're running big-time data centers, and he says that Intel's abilities shouldn't be underestimated. After acquiring various businesses that could be used to add new technologies to its microprocessors, he points out, Intel has the tools it needs to customize chips for customers. Like Andrew Feldman, however, he questions whether this makes economic sense for the company.
What's more, Ferguson says, the web giants would have to acquire the right engineering talent. But that's not a big hurdle. These companies have a relatively long history of hiring hardware engineers, and Google has already acquired a chip designer: the San José, California-based Agnilux.
With Google and Facebook designing everything from servers to storage and networking gear, you could certainly see them customizing their own chips somewhere down the road. AMD's John Williams says this would be quite an undertaking -- even for Google. But he doesn't rule it out. A company like ARM, he says, "lowers the bar" because it lets companies built their own chips without starting from scratch.
"You have to have staff to build the design. But you also have to deal with a lot of intellectual property that surrounds the processor, and at the back-end, you still have to develop a relationship with a manufacturing partner -- and that's when the costs really rise," he says. "But if you're dealing with a very significant volume, those economics could work out."
