To make server upgrades easier, Intel introduced a rack reference architecture that speeds up data throughput while reducing energy and maintenance costs in data centers.
The architecture, announced on Tuesday, calls for decoupling processors, memory and storage, and putting them in separate boxes. That is a change from the industry-standard server design, in which the processor and memory reside in a single chassis.
The reference design could potentially change system topology and reorganise traffic patterns between CPUs, memory and storage in data centers. Intel said data would move faster and in a more energy-efficient manner with the design, which could help process data and serve up results faster.
The rack-level reference architecture will be officially released next year. The architecture was introduced ahead of the Intel Developer Forum show in Beijing, which will be held on April 10 and 11.
The decoupling of units could increase the life of servers, which currently averages 18 to 24 months, and make upgrades easier, said Lisa Graff, vice president and general manager of marketing at Intel's Datacenter and Connected Systems Group.
Companies won't have to replace entire server units if a component is out of date. There will be flexibility to mix and match components in data centers depending on workloads, Graff said.
For example, companies can easily add CPUs in a hyperscale server model to deal with a growing load of private cloud or Internet transactions. Hyperscale servers are a class of systems in which computing resources can be quickly added to scale up performance. Intel expects CPU, storage and memory utilization rates to be higher with the mix-and-match model.
As Internet use grows, more data is being collected by sensors and mobile devices, and being sent to servers for processing and analysis, Graff said. There is a need for faster processors and improved server and rack designs so results can be delivered quicker, Graff said.
"The economics at this time is ripe for innovation," Graff said. "It's an opportunity for the industry."
The rack-level architecture is expected to be a followup to Project Scorpio, a rack-level reference design in which power supply, cooling and some network modules are shared. The specification was jointly defined by Alibaba, Intel, Baidu, Tencent and China Telecom.
Server products are playing an important role in Intel's operations as the company's core business of laptop and desktop chips suffers from a slowdown in the PC market. Intel is quickly expanding its data center offerings to include fabrics, interconnect and networking products, which help connect servers, storage and other components in a data center.
Server makers Dell and IBM already offer the option to decouple storage and memory with the help of specialised chips. However Intel believes there is a need for better throughput and traffic management in data centers so the components could be truly decoupled.
To that effect, Intel is building proprietary fabric that virtualizes I/O and networking to move data between servers. Intel earlier this year also said it would implement fiber optics that use light as a faster way to move data inside computers, replacing the electrical wiring.
"We've been investing in networking and interconnects ... to enable this kind of disaggregation," Graff said.
Advanced Micro Devices offers its own proprietary fabric that virtualizes I/O and storage, reducing the need for components in a server. The company has said it is looking to expand the use of that fabric, called Freedom Fabric, beyond servers.
Intel's Graff did not say whether the reference architecture had any relationship with the Facebook-backed Open Compute Project (OCP), which is trying to standardise rack and server architecture and components. OCP is working on the development of the Open Rack specification, which covers motherboards, power components and other hardware that goes inside a server. The Open Rack specification was inspired by Project Scorpio, and OCP hopes to release a standardised rack design sometime this year.
But Intel backs OCP, and has already introduced its silicon nanophotonics interconnect at an OCP event in January. Intel's Graff said the company is trying to push open standards, but did not comment on whether its rack-level architecture had any ties to the Open Rack specification.