We help our customers develop breakthrough products and services within their market.

Cambridge Consultants has been advancing research into machine learning and artificial intelligence. One area of important research has been in neural networks. These have shown impressive results for analysing real world data in the digital domain. However, the reason for their success only in the last decade is the same reason that they are difficult to use outside of the datacentre: they take an incredible amount of computation, and thus require considerable power to run.

Cambridge Consultants’ Sapphyre platform has proven itself as a powerful technology for enabling low power, high performance Digital Signal Processing (DSP) tasks, such as Software Defined Radio (SDR) and audio.

Our Sapphyre Neural Network demonstration reveals some of the research that we have been doing to slim the amount of power required to mere sips of electricity, when running on the Sapphyre platform.

For our demonstration we concentrated on an industry standard problem, image classification using the CIFAR-10 dataset. This consists of 32 x 32 pixel colour images, which are to be identified as one of ten different image categories. Check out the low-res reality in the image below:

 Two small neural network architectures were chosen that have previously been used for this task: modified versions of the VGG (Ref) network as well as the CIFAR-10 tutorial (Ref) network. This is a start: networks implemented with a typical floating point architecture are shown on the graph below as the “VGG Reference” and the “CIFAR-10 Reference” points.

The next step is to move these to the Sapphyre platform, corresponding to the “VGG” and “CIFAR-10 High Power” points. A substantial amount of energy is saved with a negligible change in final accuracy, just by moving to this efficient processing platform, with customised digital hardware extensions.

On top of the inherently low power nature of the Sapphyre system, using cutting edge algorithms allows a trade-off to be made between the final accuracy of the network and the power required. One example of a suitable algorithm is XNOR-net. By leveraging the XNOR-net algorithm for one or more layers of the network, a sliding scale of accuracy can be achieved. This is shown in the jump from “VGG” to “VGG Low Power” and the jumps from “CIFAR-10 High Power” through to “CIFAR-10 Ultra Low Power”:

Assuming a CR2032 coin cell has a capacity of 200 mAh at 3V, then 2160J of energy is available. For the lowest power network, consuming just 86 uJ of energy per classification, it is possible to run 25 million classifications from one coin cell. This opens the door to applications running for weeks from batteries, or from energy harvesting systems. The same calculation for the 7J of energy of the VGG Reference network yields an unimpressive 308 classifications.

Sapphyre is an excellent platform on which to develop these techniques. The ability to customise the architecture for the more unusual processing aspects of neural networks and XNOR-nets allows the algorithm designer the flexibility to drive the hardware design. This is explored (by yours truly) in The art of weaving algorithms into Silicon. This flexibility pushes the ability to create such an efficient system, which runs at power levels very near that of pure hardware, while remaining future-proof because it is still re-programmable.

Sapphyre’s inherently multi-core capability and it’s rich history in radio applications opens up the possibility to fuse machine learning with one of a plethora of different radio standards becoming universal in Internet of Things and Edge Compute applications.

This is just one example of Sapphyre and neural networks. This is not the only type of neural network, not the only type of machine learning and not the only type of edge compute. Different applications have different data and different performance targets, and there are a huge range of techniques that can be combined with the Sapphyre platform to deploy intelligence to places previously considered beyond reach. Our history of signal processing, chip design and machine learning means Cambridge Consultants is well placed to evaluate the individual needs of the next generation of digital processing applications, and take those through to real products.



Brendan Gillatt