How Intel plans to change servers as it breaks away from PCs

06.06.2016
From PCs to servers, Intel is trying to redesign the way computers operate. We've already seen how PCs are changing, with 2-in-1 hybrids and tiny Compute Sticks, but some of the chip maker's groundbreaking technologies will initially appear in servers.

The PC market is in decline, and the chipmaker has cut unprofitable products like smartphone chips. Intel is redirecting more resources to develop server and data-center products, which are already money makers for the company. 

Intel is also focusing on markets like the Internet of Things, memory, silicon photonics, and FPGAs (field programmable gate arrays), all of which have ties to the fast-growing data center business.

Intel has cut about 12,000 jobs in the transition away from smartphone chips and PCs. Employees have bought into the company's new strategy, Intel CEO Brian Krzanich said during a speech at the Bernstein Strategic Decisions Conference last week.

Many innovations and "dramatic changes" are coming over the next two to three years, especially on the data-center side of the business, Krzanich said.

Intel has always looked at ways to drive up performance in individual systems, but the company's focus is changing to drive improvements in server, memory, networking, and storage components at the rack level. The company is also working to speed up communications between the components.

"We have a lot of good work to do," Krzanich said.

Intel has pushed a concept called the Rack Scale architecture, which is meant to bring configuration flexibility and power efficiency to server installations. The idea is to decouple processing, memory, and storage into separate boxes on a rack. More memory, storage, and processing resources can be installed at the rack-level than on individual servers patched together, and shared resources like cooling could help cut data-center costs.

Intel's OmniPath fabric, a superfast interconnect technology, is viewed by Krzanich as the centerpiece of new server technologies. It will provide the protocols for CPUs to communicate at faster speeds with components inside a server and at the rack level. In the future, Intel envisions data transfers happening over beams of light, which will speed up OmniPath.

OmniPath will accelerate workloads like analytics and databases. It will be available through network controllers on Intel's upcoming Xeon Phi supercomputing chip code-named Knights Landing, but the ultimate goal is to bring the interconnect closer to the CPU.

"There are workloads that can be taken from software that's working in memory to software that's working right on the silicon, right next to the CPU, with a direct link through the OmniPath fabric," Krzanich said.

Using beams of light for speedy data transfers is the idea behind silicon photonics, another technology that's a priority for Intel. Silicon photonics will replace traditional copper wires, and bring faster data transfers across storage, processing, and memory components on racks, Krzanich said.

After delays, Intel has said it will ship modules to implement silicon photonics later this year.

Xeon chips will always be important to Intel, but the chipmaker is also looking at speedy co-processors called FPGAs to quickly perform specific tasks. Intel believes a killer combination of CPUs and FPGAs, which can be easily reprogrammed, could speed up a wide range of workloads.

FPGAs are already being used by Microsoft to speed up the delivery of Bing search results, and by Baidu for faster image search. Intel believes FPGAs are relevant for artificial intelligence and machine learning tasks. Intel also plans to use FPGAs in cars, robots, drones, and IoT devices.

Intel acquired FPGA technology through the US$16.7 billion purchase of Altera last year. The company's next step is to pack an FPGA alongside its Xeon E5-2600 v4 server processor on a modular chip. Ultimately, FPGAs will be integrated on server chips, though Intel hasn't provided a timeline.

Intel is also developing a new type of storage and memory called 3D Xpoint, which the chipmaker claims is 10 times denser than DRAM, and 1,000 times faster and more durable than flash storage. Krzanich described 3D Xpoint as being a "hybrid between memory and storage." The technology will first come to gaming PCs under Optane-branded SSDs, but will branch out to servers in the form of flash storage and DRAM modules.

The emerging technologies from Intel may require companies to change their server architectures from top to bottom. But as long as the servers deliver cost-performance benefits, the technologies will be adopted, Krzanich said.

Intel hasn't yet provided a cost estimate for the investment, and it hasn't described how racks infused with new technologies could be implemented alongside existing server installations. Intel will continue selling regular server CPUs, but it may take time for customers to adopt the new technologies until they are proven. 

Intel held a 99.2 market share for server processors in 2015, but that may fall next year as AMD releases new server chips and the adoption of ARM servers potentially grows.

Agam Shah

Zur Startseite