Intel to customize chips for big data applications

Intel, increasingly customizing server chips for customers, is now tuning chips for workloads in big data.

Software is becoming an important building block in chip design, and customization will help applications gather, manage and analyze data a lot quicker, said Ron Kasabian, general manager of big data solutions at Intel.

Through hardware and software improvements, the company is trying to figure out how its chips can perform better in areas like predictive analytics, cloud data collection and specific task processing. The company has already released its own distribution of Hadoop, a scalable computing environment that deals with large data sets, and now chip improvements are on tap.

Kasabian said Intel is starting with the software. "It takes a while to get silicon to market," he said. "We understand where we can optimize for silicon, and there are certain things to [improve] for performance and optimization."

The company is taking lessons from software implementations and then looking to enhance the silicon to fill any software gap, Kasabian said, adding that the chip-design process takes about two years.

Server makers have been customizing servers specifically to carry out big data workloads, and improvements at the chip and instruction-set level could speed up task execution.

The plan includes developing accelerators or cores for big-data type workloads. For example, Intel is working with Chinese company Bocom to implement the Smart City project, which tries to solve counterfeit license plate problems in China by recognizing plates, car makes and models. The project involves sending images through server gateways, and Intel is looking to fill software gaps by enhancing the silicon. One improvement could be implementing accelerators to decode video, Kasabian said.

Intel has a big software organization, and the appointment of Renee James -- formerly head of the software unit -- earlier this year as the company's president was a sign of the chip maker's intent to dig deeper into software. The company does not want to become a packaged software distributor, but wants to enable software to work better on Intel architecture hardware. For a long time, Intel has backed open-source software and has hundreds of coders contributing to the development of Linux.

Different industries have different implementations of big data, Kasabian said. For example, a big data problem in genomics could differ from one in telecommunications.

Intel is also entering the space of the Internet of things, an emerging field in which networked devices with embedded processors and sensors are used as data-gathering instruments. The company has assets such as McAfee's software and hardware platform and Wind River's real-time operating system for its embedded chips to quickly process and securely collect data.

Outside of the silicon, Intel is focusing on providing the right software tools for data centers. Hadoop was the starting point, and now Intel is looking closely at analytics, Kasabian said.

Attaching Intel's name to Hadoop will "kind of ease the mind of folks in enterprises," Kasabian said, adding that implementation of the platform in data centers will be easier.

A lot of research is also taking place at Intel labs on stream processing and graph analytics as the company designs chips and tweaks software.

"We're looking at all the big industry categories," Kasabian said.

How Intel's modded chips make your smartphone a better listener

For years, Intel and other chip makers designed processors like stock engines, dropping them into PCs, notebooks, and servers. Now, Intel has shown a newfound willingness to mod custom silicon for server customers, tweaking them with hardware and software accelerators to improve their performance.
These same servers are the ones powering cloud applications that include email and storage, but also interpret the gestures and spoken commands of smartphone users. On Monday, for example, Intel and Nuance Communications disclosed that Intel is developing an accelerator to improve Nuance’s voice recognition, which powers as many as 6 billion connected devices, according to Sean Brown, Nuance’s senior manager of innovation.

Intel has also developed custom chips for both eBay and Facebook, said Jason Waxman, general manager of Intel’s data center group, at an Intel datacenter event on Monday.

IntelIntel’s Jason Waxman holds a server motherboard.
The bottom line? By specifically improving these cloud functions, smartphone and PC users will see their performance increase over time, for free, with new capabilities. “With advanced natural language processing, you can really say anything to [connected devices] and have a natural response,” Nuance’s Brown said. “But the personal systems of tomorrow will be proactive, when you need it automatically.”

Intel has been designing “accelerators” for decades, Waxman explained, dating back to the MultiMedia eXtensions (MMX) instruction set that it shipped with 1997’s Pentium MMX. Those chips included dedicated hardware registers that were designed to accelerate software written with the MMX software in mind, a technique that Intel used with subsequent chips up to and including its Core processors. Most of today’s server chips ship with dedicated hardware logic blocks on board designed to specifically accelerate functions like encryption. Accelerators for technologies like Nuance take the same approach, but are optimized for a single company.

IntelIntel’s Pentium with MMX Technology.
Meanwhile, incoming Intel chief executive Brian Krzanich has invited non-competing chip customers to manufacture their chips at Intel’s own fabs, analyst Patrick Moorhead of Moor Insights and Strategy noted. Intel has had partners that would like Intel to design accelerators around their technology, but lacked the resources, Waxman said. The new spirit of collaboration benefits both sides.

Waxman referred to Intel’s new approach as an “SoC methodology,” an industry term for the systems-on-a-chip (SoC) approach used by many consumer manufacturers, including rival AMD, to optimize chips for specific products.

As an example of Intel’s commitment to custom chips, Intel announced a 14-nm “Broadwell SOC,” Intel’s first semi-custom chip, that will ship as part of Intel’s Atom line in 2014 or later. For years, Intel has shipped its Xeon chip to server customers, but recent shifts in the server space caused Intel to beef up its Atom tablet and set-top chip line to address enterprise customers.

In part, Intel’s hand is being forced by ARM and the smartphones and tablets that those chips power. ARM is riding three trends: first, a general shift away from the PC—Intel’s center of power—to slimmer, cheaper, tablets and smartphones. That’s a segment of the market Intel hopes to address with its Atom processor, whose “Bay Trail” derivative could power Windows and Atom tablets and convertibles as low as $150 by the holidays.

Intel has already begun mixing and matching chips to various customer segments.

An ARM chip.
”The first thing [that’s different about Intel] is that they’re using an array of products,” said Nathan Brookwood, principal analyst at Insight 64. “If you go back a few years, they designed a chip for the desktop, and then they optimized it a little bit for mobile, and they put a bigger cache on it and called it [a] server [chip]. And there were three basic markets being served by one design. And now, they’re taking stuff that was designed for mobile... Atom stuff, that can also can go into servers, with server stuff. And now they’re taking another chip that they designed for a laptop, and they’re beefing that up... for desktops.”

The second trend that benefits ARM is that ARM, unlike Intel, doesn’t manufacture its own chips. Instead, it licenses its processors as essentially a kit—intellectual property that can be used as is, or surrounded with a chip designer’s own peripheral logic. That’s forced Intel’s hand, Moorhead said. “So they’re going to have to be custom, and more flexible, to keep their 95 percent market share,” Moorhead said of Intel.

Finally, while the PC has remained generally unchanged for decades, server manufacturers have shown a willingness to optimize their servers for specific functions, known as “disaggregation”. Currently, racks of servers perch one over the other in a server chassis, all with their own CPU, storage, and memory.

 Eventually, the idea is that servers will have “pools” of computing, storage, and I/O resources, much like the a bunch of flash chips are combined together to make a single flash “drive”.

In recent years, Intel has arguably invested more in the server space than it has in the PC, into software, services, networking, storage, and other areas. For right now, the trend is toward custom-designed servers, with specific applications in mind.

“Flexibility is the key, because the architecture is disaggregating, meaning the only way to efficiency in the datacenter is workload by workload, where virtualization is not cutting it,” Moorhead said. “In the scale out data center, you have to have a specific server type for a server workload, whether it be Web tier, data mining, Hadoop or otherwise.”

The talk now is of a software-defined future, where applications “ask” for resources, which are managed by a central server—powered by Intel, the company hopes. ARM is hoping to cash in on the third trend, with low-power 64-bit server processors it will ship later this year. For right now, however, Intel hopes it can control the conversation—part of the reason executives are holed up with analysts and customers this week.