British startup Pragmatic Semiconductor has introduced a 32-bit flexible microcontroller for running machine learning models at a price of less than $1. The Flex-RV microcontroller is based on the open RISC-V architecture and uses a different material instead of traditional silicon.
Image source: nature.com
Silicon in Flex-RV is replaced by indium gallium zinc oxide (IGZO), and a key innovation is the elimination of complex and expensive packaging that fragile silicon chips require for mechanical protection. IGZO transistors are printed directly onto a plastic substrate at low temperatures. Pragmatic demonstrated how a flexible microcontroller was wrapped around a rod, and then, once unrolled, it executed code normally. At the demonstration, the company specialist handled the component very carefully, that is, the microcontroller does not promise complete freedom of action, but the available capabilities also seem promising.
You can’t expect high performance: the Flex-RV prototype has only 12,600 logic gates on board, and its clock speed is only 60 kHz, or 0.00006 GHz when compared to gaming hardware. But there is a low-power accelerator for machine learning algorithms. It is not suitable for GPT-4 training, but with a power consumption of less than 6 mW, Flex-RV is ideal for use in disposable medical devices, flexible wearables, soft robotics and brain-computer interfaces. RISC-V’s open instruction set means there are no licensing fees to pay and this helps reduce the cost of the component.
During testing, the microcontroller continued to operate when bent up to a radius of 5 mm. Throughput varied by a few percentage points, but overall the Flex-RV handled bending quite well. Its difference from many similar projects is that it uses non-rigid components on a flexible printed circuit board – the microcontroller integrated circuit can also twist and deform.