Intended to support high-performance computing, IoT appliances and wireless access networks handling 4G and higher traffic, the platform features deep learning and pattern recognition computing capabilities. By installing the hardware at multiple, geographically dispersed base stations along the edge of the network, the platform looks to distribute computing capacity and deliver quick analysis of locally generated data.
Sailesh Chittipeddi, IDT's vice president of Global Operations and chief technology officer,refers to the platform as "supercomputing at the edge" because the technology is deployed and analysis conducted at local base stations—at the edge of the wireless network—rather than in a central location, removing the bottleneck between the base station and the core of the network.
"By taking mobile low-power GPU technology and connecting it with 100 ns latency RapidIO interconnect, this modular cluster can be deployed to distribute high-performance compute functionality to the edge of the wireless network, where it is most geographically sensitive," said Jag Bolaria of the Linley Group. "The innovation paves the path for co-locating real-time deployable analytics in the approximately 2 million base stations deployed annually in wireless networks."
The Supercomputing at the Edge platform uses IDT's 20 Gbps interconnect technology to connect a low-latency cluster of NVIDIA Tegra K1 mobile processors. It's suitable for micro base station deployment along with larger computing clusters in the C-RAN, a new cellular network architecture. Each computing card is based on connecting up to 4 GPU units per processing card connected with RapidIO low-latency NIC and switching products on board.
The platform, which can support up to 12 teraflops per IU RapidIO server blade, is based on servers from Prodrive Technologies and computing cards from Concurrent Technologies PLC. Each computing card contains four NVIDIA mobile processors. The processors deliver 192 fully programmable CUDA cores for advanced graphics and compute performance. Each card in the system matches computing cores with 20 Gbps interconnect to each GPU, with over 140 Gbps of built-in switching at each node with IDT's best-in-class timing solutions