Lenovo developing ARM servers with UK Science and Technology Facilities Council

Lenovo is testing ARM-based SoCs with specialised hardware for workload-specific tasks

Lenovo is partnering with the UK's Science and Technology Facilities Council (STFC) to build and investigate the suitability of ARM-based server designs for specific workloads where high performance and energy efficiency are key concerns.

Fresh from completing the acquisition of IBM's x86 server division, Lenovo is continuing IBM's existing relationship with the STFC on research projects.

The latest collaboration will see the STFC's Hartree Centre deploy and test ARM-based server hardware from Lenovo to investigate whether it can meet the challenges faced by data centres after explosive growth in demand.

However, this is seen as a long-term investment for Lenovo, which said that it is unlikely to lead to a commercial ARM-based server offering in the short term.

Nevertheless, the move means that all three of the major server vendors - Lenovo, HP and Dell - have ARM-based systems in production or under development.

Lenovo is not looking just to stick a bunch of ARM chips into a chassis and deliver an ARM-based alternative to a Xeon server. Instead, it is building systems that will be optimised for specific kinds of workloads, some traditionally served by high-performance computing (HPC) architectures.

"Traditional servers have standardised on a common foundation, then you stack on top different kinds of cards that give the server its personality for a certain workload, such as networking cards, security accelerators, field-programmable gate arrays or GPUs," Lenovo's executive director for HPC, Scott Tease, told V3.

The problem with this approach is that the cost adds up, as it bumps up power consumption and often adds latency to applications, according to Lenovo.

"So what we're trying to do is figure out if there is a better way to go, where we look at more workload-optimised systems where all the functionality is designed into the base of the system," Tease explained.

To this end, Lenovo is using a ThunderX system on a chip (SoC) from Cavium. This boasts 48 ARMv8 64-bit cores, and Lenovo is fitting two into each node to deliver a 96-core server.

The nodes are designed to fit into the same enclosure as the NextScale systems that IBM launched in 2013. With 12 nodes per chassis, this means that the system can fit 1,152 cores into 6U of rack space.

As a SoC, the ThunderX has on-chip accelerators for functions such as compression. It also features four built-in 40Gbps Ethernet connections, which can be broken out as 16 10Gbps connections and used with a virtual switch capability as a fabric for high-speed connectivity between nodes.

"What you can do with that is connect the nodes together in a mesh topology that allows for node-to-node communication without the need for any switching at the top of the rack," Tease explained.

Lenovo and the STFC are looking at applications including cloud, search and web serving and caching, plus HPC.

"The reason we selected these is because they all have similar technical requirements, they are all looking for good energy efficiency, and users are looking for the best cost for the level of performance," Tease said.

In line with other ARM server customers, such as those using HP's Moonshot platform, Lenovo said that potential users have a strong desire for open systems such as Linux, and many are running their own code stack instead of off-the-shelf software. They are also running at large scale so they can afford a workload-specific server optimised for their requirements.

"There have been a lot of announcements about ARM servers, but this is different. This is not a lightweight ARM, like a mobile phone, this is a higher performance, real server kind of ARM architecture," Tease said.

Lenovo does not see these systems becoming generally available anytime soon, but the firm disclosed that it is working with selected customers, which it declined to name, to test the viability of the ARM architecture for specific workloads.

"Hartree will be the first to get these systems, and will be focusing on the energy efficiency side of this and also the ecosystem building side. We have other partners also in the EU looking at GPU attachment or utilising the network efficiency," Tease said.

"We want to stay ahead of whatever emerges as we look to take this from proof of concept to real deployment."