Imagination unveils neural network accelerator

1 min read

Imagination Technologies has unveiled a complete standalone hardware IP neural network accelerator (NNA) using a specialised PowerVR architecture that’s able to support high performance computation.

NNAs are a fundamental new class of processor and will, according to Chris Longstaff, Imagination’s senior director of product and technology marketing, PowerVR, ‘become a standard IP block on future SoCs just as CPUs and GPUs have done’.

Potential applications for NNAs include: photography enhancement and predictive text enhancement; feature detection and eye tracking in AR/VR headsets; automotive safety systems; facial and speech recognition; and collision avoidance and subject tracking in drones.

As neural networks become increasingly common, dedicated hardware solutions like Imagination’s PowerVR 2NX NNA – said to provide an 8x performance density improvement versus DSP only solutions – will be required.

Neural networks are bandwidth hungry and the 2NX has been designed to minimise bandwidth requirements for the external DDR memory to ensure a system is not bandwidth limited.

The 2NX includes hardware IP, software and tools to provide a complete neural network solution for SoCs. Depending on the computation requirements of the inference tasks, it can be used standalone – with no additional hardware required – or in combination with other processors such as CPUs and GPUs.

Imagination has also launched a new generation of PowerVR GPUs.Building on the company’s PowerVR Rogue architecture, the Series9XE and 9XM GPUs continue to extend the fillrate/mm2 of the 8XE series of GPUs. The range has a new eight pixel/clock core, suitable for GUI focused applications requiring support for 4K60.

The Series9XM GPUs also benefit from improvements in the memory subsystem, reducing bandwidth by as much as 25% over previous generations to ensure the increased processing power can be fully used.

Other features include a new MMU, allowing a greater address range and standard support for 10bit YUV across the range, without impacting area.