Low-latency, local computing speeds big data analytics

October 15, 2015 // By Graham Prophet
From IDT’s Open High-Performance Analytics and Computing Lab, this platform is optimised for computing services in access networks and video network distribution: it facilitates “deep learning at the edge of networks”.

Integrated Device Technology has launched a heterogeneous mobile edge computing platform that performs real-time Big Data analytics and deep learning with low-latency computing at the network edge. Developed at IDT’s Open High-Performance Analytics and Computing (HPAC) lab, the platform uses modules currently in production and is based on innovations and collaboration with NVIDIA, Prodrive Technologies (Eindhoven, Netherlands) and Concurrent Technologies PLC (Colchester, UK).

“Network operators are looking for ways to add analytics capabilities on large amounts of unstructured data close to users, by pushing more compute capacity to the network edge, rather than sending data on the backhaul to the core of the network data centres,” said Sailesh Chittipeddi, chief technology officer and vice president of Global Operations at IDT. “IDT’s Mobile Edge Computing Platform offers network and data centre operators, as well as systems integrators, a quick path to computing at the edge by connecting accelerators based on a heterogeneous architecture with IDT’s low-latency RapidIO products. “This is becoming more important with high-bandwidth streaming video and associated analytics that need to happen close to the last mile where users are located,” Chittipeddi noted.

The new platform is based on a 1U form factor server and features heterogeneous computing with NVIDIA Tegra K1 processors and x86 CPUs. The computing elements are connected using RapidIO fabric technology with about 100 nsec switching latency and 20 Gbps ports, with forthcoming versions using RapidIO 10xN technology accelerated to 50 Gbps. The platform can be populated with appropriate mixes of processor types and accelerator capacity to optimise workloads for deep learning and analytics at the edge of the wireless network. Co-located with base stations, it is possible to support in ATCA form factors if required. The modular building blocks can be deployed by carriers for server functionality at the wireless base station, Cloud Radio Access Network (C-RAN) or in the central office. And with distributed data centre functions becoming increasingly important for streaming