STMicroelectronics has announced a new type of chip with the launch of the intelligent sensor processing unit (ISPU) that combines a MEMS sensor and digital signal processor (DSP) that can run AI algorithms ON the same silicon. The AI sensor is said to provide benefits in four areas: power consumption, packaging, performance, and price.
By combining the signal processing and AI algorithms on MEMS sensors, it saves space and reduces power consumption by up to 80%, which also cuts costs, said the company. ST calculates a 5-6× power savings over system-in-package approaches in sensor fusion applications, and a 2-3× savings in run mode.
The launch of the ISPU also debuts ST’s concept of “Onlife Era” in which smart sensors can “sense, process, and take actions.” It also migrates intelligent processing from “at the edge” to “in the edge,” said the company, and enhances privacy by keep the data local.
The proprietary ultra-low-power DSP, an enhanced 32-bit reduced instruction set computing (RISC) machine, can be programmed in C. It also allows quantized AI sensors to support full- to single-bit-precision neural networks for higher accuracy and efficiency for tasks such as activity recognition and anomaly detection by analyzing inertial data.
The DSP is extensible (in the chip-design phase) for dedicated instructions and hardware components. Other processor features include full-precision floating-point unit, a fast four-stage pipeline, 16-bit variable-length instructions, and a single-cycle 16-bit multiplier. Interrupt response is four cycles versus 15 (Cortex). The frequency/ODR max is 5 MHz/3.33 kHz –10 MHz/6.66 kHz.
ST’s sensors with ISPUs will be housed in 3 × 2.5 × 0.83 mm packages and will be pin compatible with predecessors for easy upgrades. The ISPUs are programmable with AI commercial models and interoperates with Keras, TensorFlow Lite, ONNX, and Scikit-learn. The ISPU is available with NanoEdge AI for self-learning to help reduce development time.