We’re delighted that the design team for our PowerVR Series2NX Neural Network Accelerator (NNA) has been honoured with a prestigious British Engineering Excellence Award (BEEA). The BEEAs were established in 2009 to demonstrate the high calibre of engineering design and innovation in the UK. This year 40 different companies competed across 11 categories. The NNA design team received a BEEA for ‘Design Team of the Year.’
This is actually the second time this year that PowerVR designers have received an award for their work on the Series2NX – in May, the PowerVR Vision and AI team were named ‘Tech Team of the Year’ at the National Technology Awards 2018.
So, what’s so special about what this team has accomplished?
It takes a hard-working team to deliver a leading product and all the required software tools from the ground up in just 14 short months.
The project began in 2016 with a small dedicated group of engineers and limited resources. A mix of research, hardware and software engineers, they saw a need for the product and set out to make it a reality. As management saw initial results, a larger focus was put on the product, and the team grew steadily to realise and implement requirements. Ultimately, the team behind the Series2NX, led by Imagination’s VP of Vision and AI Russell James, comprised 90 engineers across four countries.
At the concept stage, industry solutions for computer vision were not fast or efficient enough to be able to execute the compute-intensive operations required for neural network acceleration on edge devices.
Given its deep experience in designing leading GPUs for mobile and embedded markets, the design team initially started the Series2NX design by modifying the PowerVR GPU architecture to achieve the performance levels needed for neural network acceleration. However, this approach led to a solution that was too large, and that idea was put aside. The team decided to create a completely new design. The Series2NX was created as a specific, optimised hardware acceleration unit for convolutional neural networks.
Once they made the decision to create a new design, the target was to achieve ~10-100x higher performance per mW and performance per mm2 compared to other embedded NNA solutions in the market. The end result exceeded expectations. By October 2017, the first evaluation systems were released, and the production release of the initial Series2NX IP, the PowerVR AX2185, was delivered soon thereafter.
Designed from the Ground Up
Every feature of the Series2NX was designed specifically for it. There were no preexisting flows, tools or system architecture to help ease or speed up the process. Of course, because everything needed to be custom designed, it presented the team with a unique opportunity to create the most efficient and competitive product they could without any constraints – something the team embraced and managed to do with great success.
One challenge was in creating an infrastructure to validate the product and develop all the tools. This was a complex task because each different network type has different testing needs – for example, a system for recognising faces has very different needs from a system for interpreting sentences.
The team came up with a comprehensive combination of tools, and supporting databases, for both bit-accurate and for fuzzy comparisons between the various components of the system. The hardware, simulation, software and quality assurance teams worked together finding practical solutions with a lot of trial and error, as well as using more theoretical techniques from the research and architecture teams.
The Series2NX NNA Today
Today, we offer two Series2NX cores – the PowerVR AX2185, optimised for performance efficiency, and the PowerVR AX2145, and multiple licensees using it for a range of applications including mobile and automotive. We’re excited about the traction it is gaining in the market, the amazing products our customers are rolling out, and the potential for future development, based on the innovations of our incredible PowerVR Vision and AI team.