ARM’s next bet for plastic chips: Neural networks
Back in 2015, ARM’s CTO Mike Muller had wowed ARM TechCon’s attendees with a fully flexible ARM1 SoC designed at a 3µm node on plastic, sporting about 25k transistors.
Talking about the progress realized since then, White noted that after a low-end Cortex-M SoC iteration with a footprint of roughly 8 cm2, PragmatIC’s latest layout of the SoC had shrunk further to 1cm2, and the company is now moving to a new plastic process node with new design rules which the CEO is confident will further reduce the device’s footprint.
Referring to the process node roadmap flashed by Muller at ARM’s TechCon, White says PragmatIC is in line with the tentative plastic roadmap, roughly halving the devices’ footprint every year, somewhat faster than the equivalent Moore’s law for silicon.
But the CEO doesn’t expect the Moore’s law analogy to last very long. “We’ll reach a plateau where our technology will be suited for the realization of lower-end MCUs or below. Driving the Cortex-M project was more of an R&D challenge to see how far we could push the technology, but there are very few practical packaging applications today that would require a full 32-bit SoC and where it would make economic sense”.
“The types of circuits we aim for are in the region of a few thousand gates to add intelligence into high volume fast moving consumer goods, then in the next few years 10,000 gates may become a sweet spot for us, in applications where flexibility is a key product benefit” White added.
But MCUs and conventional SoCs may not necessarily be the workhorses of plastic electronics as they have been in the silicon industry. ARM is an investor in PragmatIC and is also the lead development partner on the PlasticARMPit project initiated last October together with consumer goods giant Unilever and the University of Manchester. In the Innovate UK-funded project description, the partners aim for the design of highly energy-efficient processing engines for flexible sensors, targeting specific sensor data for which plain flexible MCUs are unlikely to meet the necessary computational needs.
As of digital processing engines, the partners propose to develop plastic Neural Networks (NNs) customised for specific applications and capable of operating in extremely parallel fashion to achieve high performance at low power. With this project, both ARM and PragmatIC hope to establish digital hardware NNs as the de-facto processing engine for printed electronics.
“NNs are particularly interesting for sensing applications in the real world, with a combination of different sensor inputs. They are good at classifying the data so it can be interpreted according to the categories of results you are looking for” commented White about the PlasticARMPit project.
“What’s more, the physical structure of the flexible ICs translates well into neural networks. The comparative work performance and yield of plastic electronics is less of an issue in neural networks where there is a lot of redundancy” added White, saying that imprinted electronics on plastic foils allows for the modelling of large number of neurons. “Building circuits layer by layer, we can physically build something that mimics a neural network architecture. It is not for high-end machine learning of course, but for smart packaging and sensing applications, you can get a form of categorization from thin flexible circuits”.
In the case of the PlasticARMPit project, Unilever was the one who came with the business case. The idea is to couple a flexible and multi-analyte e-Nose sensor with a plastic NN onto a wearable patch to detect armpit malodour composition and determine how effective the company’s antiperspirants and deodorants are. In this particular research project running until March 2020, PragmatIC will integrate organic TFT bio-sensors developed at the University of Manchester, but outside of this project, it could take sensors from different partners.
“Our focus is on interpreting the analogue inputs and convert it to useful data. Sensing companies interested in the computing backend could come to us”, said the CEO.
After an initial market pull for so-called personal care, White hopes such neural networks on plastic could break into the health industry, since the underlining technology is applicable to many biomedical applications. With neural networks, you could have an array of different analytes and fused sensor data which would not necessarily give a predetermined 1/1 match answer but more complex diagnostics, Whites anticipates.
Regarding the IP and design approach, PragmatIC says the FlexICs can be designed using a conventional EDA flow. For now, the company is doing full custom designs as well as sharing its Process Development Kit with selected close partners (such as ARM), but in the future PragmatIC expects plastic chip design to follow the path of silicon, with standard IP libraries and third-party designs based on a PDK.
in Sedgefield, UK.
The self-contained and fully automated FlexLogIC system is modular in construction, calling for a capital investment several orders of magnitude smaller than a new silicon IC fab, yet opening up the potential for a distributed and highly scalable manufacturing model for embedding electronics in everyday objects. Production cycle time is under a day compared with over a month for a silicon IC, making it possible to develop and test flexible electronic solutions at very short notice before mass deployment.
PragmatIC is now focused on ramping its production in order to fulfil demand for applications with some of the worlds’ largest consumer brands. That includes partner and investor Avery Dennison, the world’s leader in RFID labels committed to deliver fully printed RFID labels for specific consumer applications.
ARM – www.arm.com
PragmatIC – www.pragmatic.tech
Related articles:
Personalized cosmetics: a low-hanging fruit for printed biosensors
Pragmatic’s “fab-in-a-box” to beat expectations
Starting all over again on plastic: ARM
ARM backs printed electronics startup
Isorg raises funds to ramp printed sensor production
If you enjoyed this article, you will like the following ones: don't miss them by subscribing to :
