Intel is putting a sharp focus on expanding its custom processor and chip operations in response to a growing trend of companies building servers in-house to meet specific workloads or data center designs.
Last year, Intel has supplied custom processors to 18 companies, most notably eBay and Facebook, said Diane Bryant, senior vice president and general manager of Intel’s Datacenter and Connected Systems Group, in an interview.
“That trend is growing. In the last year we’ve delivered 18 custom silicon processor solutions for the full array of customers—our direct customers, the OEMS and the end users—in order to meet their specific needs,” said Bryant, who runs Intel’s most profitable group.
The growing custom processor business will supplement Intel’s bread and butter server chip business, consisting of selling generic Xeon processors available in rack, tower and blade servers sold by companies like Dell and Hewlett-Packard. But server infrastructures are changing with the growing adoption of cloud computing, big data and other applications, which has translated into a growing demand for custom processors, Bryant said.
Companies like Facebook and Google with mega data centers design servers in-house, and get them made from direct server suppliers like Quanta. The barebones servers typically cut excess components and are good enough to handle the growing cloud transactions such as search requests and social networking tasks. Facebook and Google have experimented with ARM processors in servers, while Tilera processors have also been tested.
“When you work with these end users who have technology as their core business, they are very clear on what is required,” Bryant said. “They know what their workloads are, what their various applications are, they know what metric they are looking to hit from a performance per total cost of ownership.”
The level of processor and chip customization varies with the workload, data center design, and even cooling solutions. Bryant provided an example where a flexible cooling system in a data center would allow customers to run processors at a higher frequency.
“We will have customers that have a very [specific] power target, so we will create versions whether its through changes in frequency, changes in core count, changes to drive down the power,” Bryant said.
Customers usually give information about the applications they are running, the accelerators they need, the performance and power consumption levels they are looking to hit. Intel then customizes processors and chips that meets the specifications. Some customers in the technology and the data center business get specific about the server infrastructure.
“It all boils down to scale. I had one cloud service provider who had told me a single application is running across tens of thousands of servers. You can afford to tune that server very targeted against that application and eke out every bit of performance at ever lower cost of operations,” Bryant said.
Intel is also now able to build system-on chip (SoC) designs, in which the CPU is combined with other accelerators, I/O, graphics and other processing units. That makes it easier to build custom processors and chips, Bryant said.
“With our SoC capability now, we can actually do rapid turns of our base product with very unique accelerators. Whether it’s voice recognition acceleration or encryption or graphics acceleration… all the different types of accelerators that are targeted at different apps. We can deliver unique products there too,” Bryant said.
Prepareing for the next generation of processors
Intel next year will release Xeon server chips based on the Broadwell processor core, which will succeed Haswell. Bryant said that the server SOC will also help optimize the chip to workloads, be it analytics or cloud.
“We have this wonderful Xeon core, and now Intel has a system-on-a-chip capability where we can rapidly turn out grabbing different intellectual property blocks and accelerators. Why not take this Xeon core and marry it with the SoC capability, and come up with… very [specific] processors targeted at unique capabilities,” Bryant said.
In some ways, Intel is taking the same route as Advanced Micro Devices, which is creating custom chips based on its CPU and graphics architectures, but largely for non-server products. AMD’s custom chips will be used in the upcoming Sony PlayStation 4 and Microsoft Xbox One gaming consoles.
Intel is also investing in software development to tie applications directly to chip development. The chip maker has released its own version of Hadoop, and is also actively contributing an orchestration layer to OpenStack so resources are effectively allocated at server, storage and network levels in distributed computing environments.
Beyond the server, Intel is also looking to change data center design. One of the projects called Rack Scale aims at decoupling the processing, I/O and storage units in data centers with faster throughput mechanisms.
“Instead of a rack being 24 servers slotted in, with each of those servers with compute, memory and I/O, instead break that artificial barrier of the server down and look at it at the rack level. And create pools of compute, memory and I/O so that the application can access and use whatever capacity it needs,” Bryant said.
The company is expected to announced a new optical throughput standard called MXC, which will be detailed at the Intel Developer Forum next month. The company is also developing processors for different target markets, Bryant said. Intel will also announce a new Atom processor called Rangeley for embedded networking devices in early September, ahead of IDF.
“We have hundreds of microprocessor products to cover the entire space,” Bryant said.