Quad core IP for edge AI data processing

The SFA-200 from Sondrel combines a quad ARM core block with AI core and security sub-system with another ARM core for edge AI data processing

UK ASIC design house Sondrel has launched the latest in its series of semi-custom chip for edge AI and IoT systems.

The SFA 200 is a quad core ARM chip with added security for remote gathering and processing of video and data at the edge. The core can be used for a scalable single channel ASIC where custom modules can be added for smart metering, smart homes, smart factories, voice-controlled devices and infotainment.

Each ASIC is based on a CPU cluster for general purpose tasks and an AI cluster for neural network tasks with supporting memory and interfaces. The cluster of four ARM A53 or A55 cores can be combined with an AI Ethos core linked by an 800MHz Network on Chip. This has a data path that can be between 64bit to 512bit wide, optimised for data flow with quality of service (QoS) arbitration and scheduling support. A System Management Subsystem (SMS) manages the configuration, start-up (Boot) and operation modes of the SoC along with active power management for low or battery powered applications.

Related articles

The CPUs cores provide 9 to 10 GMIPS at 1 GHz while the Ethos AI core provides up to 4TOPS. Custom AI cores can be integrated as an alternative if required. An integral security subsystem uses either an ARM A53 or M33 to provide activity/intrusion monitoring, software signing and crypto, and authentication to resist hacking, which is further enhanced by watchman behaviour tracking and deviation detection.

This semi-custom design can reduce reduce design costs by up to 30 percent.