Inside Intel's secret super-chips: If you've got the millions, it's got the magic
Just what the heck are they building in there? We asked Facebook
Analysis This year, Intel announced a strategic change in its philosophy: if you show it a mountain of cash, it will ship you chips with customized processor instructions or altered clock speeds.
Bespoke silicon for the data center is Intel's answer to easily configurable, custom-built ARM-compatible chips: electronics giants are able to license ARM's processor core blueprints, bolt on a touchscreen controller, some wireless networking or perhaps even something suited for a server warehouse, and fabricate the component as required. And obviously, if it's just you making and using the custom system-on-chips, no one else needs to know about how the silicon actually works – a frustration for low-level third-party engineers.
Until this year, that's not been the way Intel has operated its chip-baking business; everyone could expect a level, well-documented playing field with plenty of separate chipsets out there to hook up hardware. Now its newfound customization scheme has made Chipzilla reconsider how it works with some of its largest CPU-gobbling customers.
Though chip giants have in the past given some of the choicest processors from their fabs – ones capable of sustaining greater clock speeds due to better thermal limit properties – to major customers, process improvements have evened out the variabilities among modern chips of the same design.
To give big spenders such as Facebook and eBay special benefits, Intel has had to start inserting extra – and frequently undocumented – features into its processors. This approach is sometimes called "the golden screwdriver", and can be vexing to companies outside the magic circle of firms with access to the instructions, as a fraction of the silicon they've paid for is blocked off from them.
Earlier this year Intel told us it was etching different features or instructions onto silicon for some customers. The company has also created completely new SKUs with different power or performance.
"A lot of the stuff to date has been modifying different parameters – power," Jason Waxman, general manager of Intel's cloud infrastructure group, told us this month. "We're now seeing requests that are coming in where people want to include their own instructions."
Some of these requests will give Intel "a broader portfolio on customization," he said, and could include anything from new intellectual property (IP) to features that require additional or modified on-chip structures. "The range has certainly broadened," he said.
'We're working with Nuance publicly on voice acceleration'
The company's customers can, for example, ask for alternative ways to interface or directly attach electronics to a CPU's cores, said Waxman. "As we move to a system-on-chip (SoC) methodology we have the tools to do permutations [and are] working on putting together a broader portfolio on customization," he added.
Some of this will extend to the repackaging of x86 silicon with other bits of circuitry and hardware controllers to suit the chip's application, or altering the performance of a device.
On the more basic end of the scale, "some people say, I just want more IO ports, I want different IO, I want ways of connecting and driving [data]. Could be IO for storage, could be desire for different interconnects, could be faster or low latency bandwidths. Certain accelerators represent great collaboration opportunities - we're working with Nuance publicly on voice acceleration."
To find out more about what customers are doing with Intel's scheme we spoke with one of Intel's major buyers of custom chips: Facebook.
"To the extent that we can profile our workload, being able to request specific core counts, clock rates, turbo settings for key state management - those are the kind of things we are working with Intel on," Frank Frankovsky, the company's hardware design and supply chain chief, told The Register.
"All aspects are about maximizing the useful work the CPU can do per watt per dollar," Frankovsky says. "Most of our work has been around optimizing the performance we can get per watt out of each CPU."
Some of the changes Facebook has requested from Intel include alteration to the CPU clock rate, or changes in power management or how the power-states are managed.
"We have not presented any unique IP or incorporated our IP into Intel chips," Frankovsky said. Facebook is unlikely to go to the trouble of drawing its own silicon blueprints as long as Intel pays attention to some of the social network's requests, he indicated. "I guess the way I think about it, is if at some point we're unable to influence the product, then we would be required to go a little deeper."
The majority of Facebook's custom chips are being used in the social network's main web servers, which deal with the real-time construction of dynamic pages, he said. But the company is planning to use more of Intel's customization capabilities in the future, especially if a couple of Chipzilla's experimental tweaks prove fruitful.
eBay declined to be interviewed for this piece, but did tell us the online bazaar "is working with Intel on a custom chip processor and we've been encouraged by the results to date".
To the cache and beyond
Intel is busily working to expand the types of alterations it can make for sufficiently big clients, and it may even start fiddling with the basic design of its CPUs, Diane Bryant, Intel's general manager for data centers and connected systems, told us.
This year, the company hopes to have 18 custom CPUs out to market, and thinks it can broaden that as it transitions to more system-on-chip designs. It is also trying to penetrate deeper into the core of its chips to customize it more.
"Our goal, though, is to modify things like the cache," Bryant says. "The core itself which includes the cache hierarchy - that's relatively fixed. That's a future capability that's coming."
This development may take a while, as Intel is having trouble getting enough "throughput" on the modified caches, Bryant said.
When we told Facebook's Frankovsky about the caching scheme, he perked up. "Part of the workload profiling we do gives us really good insight into how we're exercising the cache and how valuable the pre-fetch algorithms are," he said. "Being able to modify the cache hierarchy would be a really interesting thing to do. In other cases it would be cost prohibitive to build a level-one cache the size we needed." (Hear that, Intel? There's money on the table over here.)
With Intel promising more than a dozen custom designs this year, and more as it moves to an increasingly system-on-a-chip approach, the company hopes to flood the world with its bespoke chips.
"You start early and then people start to hear more," Waxman said. "They're not just the Model T. People are realizing that particularly for large-scale data centers ... this is becoming almost a focused project, it's a priority, it's becoming the new normal for our largest customers." ®
Sponsored: Benefits from the lessons learned in HPC