Imagination builds standalone hardware IP neural network accelerator

Based on a specialised PowerVR

Imagination Technologies has been showing off its standalone hardware IP neural network accelerator which is powered by its PowerVR architecture implementation for neural networks.

The tech means that companies can build SoCs for mobile, surveillance, automotive and consumer systems can integrate the new PowerVR Series2NX Neural Network Accelerator (NNA) for high-performance computation of neural networks at very low power consumption in minimal silicon area.

Neural networks such as Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs) and Long Short Term Memory networks (LSTMs) are enabling an explosion in technological progress across industries. Neural Network Accelerators are a fundamental class of processors, likely to be as significant as CPUs and GPUs, both of which Imagination already delivers.

They could end up being used for photography enhancement and predictive text enhancement in mobile devices; feature detection and eye tracking in AR/VR headsets; pedestrian detection and driver alertness monitoring in automotive safety systems; facial recognition and crowd behavior analysis in smart surveillance; online fraud detection, content advice, and predictive UX; speech recognition and response in virtual assistants; and collision avoidance and subject tracking in drones.

There is a serious need for this particular type of technology.

According to the January 2017 Embedded Vision Developer Survey conducted by the Embedded Vision Alliance, 79 per cent of respondents said they were already using or were planning to use neural networks to perform computer vision s in their products or services.

A broader range of companies will be able to develop products and services with neural networks. Imagination customers are already developing and deploying NN based systems into markets including security, mobile, automotive and a set-top box.

Jeff Bier, founder of the Embedded Vision Alliance, says: “Numerous system and application developers are adopting deep neural network algorithms to bring new perceptual capabilities to their products.

In many cases, a key challenge is providing sufficient processing performance for these demanding algorithms while meeting strict product cost and power consumption constraints. Specialized processors like the PowerVR 2NX NNA, designed specifically for neural network algorithms, will enable deployment of these powerful algorithms in many new applications.”

As neural networks become increasingly common, dedicated hardware solutions like the 2NX NNA – which provides an 8x performance density improvement versus DSP-only solutions – will be required to achieve the highest possible performance with the lowest possible power and cost. In addition, neural networks are traditionally very bandwidth hungry, and the memory bandwidth requirements grow with the increase in size of neural network models.

This introduces significant challenges for SoC designers and OEMs in designing a system that can provide the required bandwidth to the NNA. The PowerVR 2NX can minimize bandwidth requirements for the external DDR memory to ensure a system is not bandwidth limited in terms of performance. Widespread availability of dedicated hardware like the PowerVR 2NX NNA will allow for further development of applications based on these neural network technologies.

PowerVR 2NX was designed from the ground-up and Imagination says that it can provide the highest inference/mW IP cores to deliver the lowest power consumption. It can also manage the industry’s highest inference/mm2 IP cores to enable the most cost-effective solutions.

It runs on a low bandwidth with support for fully flexible bit depth for weights and data including low bandwidth modes down to 4-bit and can manage a performance of 2048 MACs/cycle in a single core, with the ability to go to higher levels with multi core.

Chris Longstaff, senior director of product and technology marketing, PowerVR, at Imagination, says: “Dedicated hardware for neural network acceleration will become a standard IP block on future SoCs just as CPUs and GPUs have done. We are excited to bring to market the first full hardware accelerator to completely support a flexible approach to precision, enabling neural networks to be executed in the lowest power and bandwidth, whilst offering absolute performance and performance per mm2 that outstrips competing solutions. The tools we provide will enable developers to get their networks up and running very quickly for a fast path to revenue.”

The PowerVR 2NX NNA is available for licensing now.