UK chip design house Sondrel has launched a semicustom ASIC for machine learning in battery powered edge applications in the Internet of Things (IoT)
The company has developed a framework of blocks that are pre-verified to speed up complex chip designs. The SFA 100 IP reference platform includes an ARM Cortex M55 processing core to handle protocols for a wired or wireless connection, with security is built in using standard secure/encrypted protocols using TrustZone and CryptoCell support. This is integrated with the ARM Corstone-300 subsystem that ensures a very high level of security.
The design includes the ARM Ethos-U55 Machine Learning (ML) processor to handle applications such as voice activation, image classification, gesture recognition, filtering, inference and tracking. If these need more memory, it can easily be scaled up via the DRAM interface.
The design has GPIO, I2C, UART and QSPI peripheral interfaces to capture sensor data, such as video, image, sound and both static and dynamic data, along with a 12bitbit audio DAC for voice commands. The low overall power requirement of the SFA 100 design enables it to operate on battery power for long periods which is complemented by low power, wireless connectivity using either Bluetooth BT5.1-LE or ZigBee.
“Knowledge is power but raw data needs to be processed to make useful,” said Graham Curren, CEO of Sondrel. “The Internet of Things is all about gathering information about what is happening in a smart house, a smart factory or a smart office to make them smart. The SFA 100 IP platform has been designed specifically for this purpose by providing a compact, powerful, compute capability to intelligently transform data into knowledge and add ‘smarts’,” he said.
“Most of the design work has already been done to create this reference design so that all that is needed is the customer’s own IP and any minor customisation