Could 'Supercomputing at the Edge' provide a scalable platform for new mobile services?

4 min read

Imagine a future in which every mobile base station is capable of instantly processing data as it is being uploaded or downloaded; where some types of data may be hugely valuable for just a matter of seconds, but which don't have to be stored in or sent to the cloud; or of a computing platform, located at the very edge of the network, where data is collected and the caching of content is done locally. All of these ideas are at the heart of the 'Supercomputing at the Edge' concept. For its proponents, it heralds a new age in big data management capable of supporting many new services and applications.

"You could, for example, tweet that you've just seen a movie and are headed out to dinner. In response, some nearby dining options could pop up on your screen," suggested Sailesh Chittipeddi, IDT's vice president of Global Operations and chief technology officer. "Or, if you are standing at a bus stop, you can use the technology to check your phone for the precise location of your bus. The possibilities are endless."

According to research conducted by the Linley Group, more than 2million base stations are installed around the world each year. If more intelligent computing could be co-located within these stations, the benefits to mobile users and operators alike could be immeasurable.

For operators, reducing the bottle neck in what is described as the network's 'middle mile', that is between the base station and data centre, together with the deployment of locally based real time application driven analytics could enable applications to react faster and allow for the delivery of more tailored services that would, in all probability, be of more value to users in a specific location – and all done in a more time sensitive way.

"By integrating a large volume of low power GPUs in a server rack at scale, 'Supercomputing at the Edge' can deliver a clear route to massive cloud based clusters capable of supporting both data intensive analytics and gaming," explained Devashish Paul, principal product manager at IDT. Speaking to New Electronics from this year's Mobile World Congress in Barcelona, Paul added: "By developing large GPU clusters with low latency and massive scalability, we will be able to deliver remarkable levels of computing horsepower to the edge of the RAN."

It goes without saying that the sending of data across networks costs time and money.

According to Paul, "'Supercomputing at the Edge' is, in many respects, being driven by the demands of the operators. At present, most applications are using the Cloud. Mobile phones access the Internet via a 3G or 4G network and data travels from the access network over the core, via the 'middle mile', to a data centre where the applications may be running in the Cloud. The necessary computing is done there and the results are then returned to the client's hardware, whether that is a phone, tablet or computer. It can be slow and data can be lost.

"In addition to managing the network proper, millions of new base stations have to be connected each year as operators respond to the explosive growth in smartphone usage. All of this is putting additional pressure on operators who have to invest in maintaining and extending additional, and expensive, capacity.

"So why not move that computing resource to the network edge? We will be able to install new hardware at numerous locations dispersed among base stations which will be able to handle big data at the local level. There will be huge potential for operators to make their pipes more intelligent and the data centres more efficient."

In February 2015, IDT, Nvidia and Orange Silicon Valley, the research division of phone operator Orange, announced they had developed a supercomputing platform capable of analysing 4G to 5G base station bandwidth data in real time.

The platform combines IDT's RapidIO interconnect technology with NVidia's Mobile Tegra K1 GPU technology. It uses RapidIO enabled servers from Prodrive Technologies and Concurrent Technologies' GPU cards with embedded RapidIO interconnect.

"Managing data locally has a number of significant advantages. Mobile users will be able to get access to real time network information which will be able to offer context related services," said Paul. "'Supercomputing at the Edge' will offer a service environment that will be able to provide not only proximity in a specific locality, but also ultra low latency and access to much higher bandwidth."

Paul believes that distributing data centre functionality comes with a significant upside for operators.

"The cost of computing services is high, but if we can move the computing functionality away from data centres so that it is managed by the operators themselves, they might be able to charge application developers significantly more as their applications will be run much closer to the end user.

"Cloud based data centres consume massive quantities of energy and require a lot of computing capacity. Because the technology we're talking about is deployed at the edge, it also removes the bottleneck that exists between the base station and the core network and it is very low power.

"This new platform is intended to support high performance computing, IoT applications and wireless access networks. It will also be able to provide deep learning and pattern recognition capabilities and benefit those companies managing data centres, it will allow them to manage them more intelligently, optimising software and how the applications themselves are run."

By taking mobile low power GPU technology and connecting it with 100ns latency RapidIO interconnect technology, the supercomputing platform can be used to distribute high performance computing functionality across the entire network.

The platform uses IDT's 20Gbit/s interconnect technology to connect a low latency cluster of Nvida Tegra K1 mobile processors suitable for very dense, low energy computing.

Each computing card in the platform is based on connecting up to four GPU units per processing card which are then connected with RapidIO low latency network interface cards and switching products on the board.

The platform has been designed to support up to 12Tflops per IU RapidIO server blade, each containing four Nvidia mobile processors. Each processor can, in turn, deliver 192 fully programmable CUDA cores to support advanced graphics and the compute performance that supercomputing at the edge requires.

"What we have created is a scalable supercomputing topology that can go anywhere in the network – wherever the operator wants to push out the compute functionality," explained Paul. "It is not only suitable for pico base station deployment, for example, but also for much larger ones, such as the C-RAN cellular network architecture being rolled out in the Far East.

"What we have created is a platform with technology that is available today and which could go into mainstream production tomorrow," he concluded.