Eta Compute has developed a high-performance ASIC and new artificial intelligence (AI) software program based on neural networks to resolve the issues of edge and cellular units without using cloud resources.
Future cellular devices that are constantly active within the IoT ecosystem need a disruptive solution that gives processing power to allow machine intelligence with low power consumption for applications like speech recognition and imaging.
These are the types of purposes for which Eta Compute designed its ECM3531.
The IC relies on the ARM Cortex-M3 and NXP Coolflux DSP processors. It makes use of a tightly built-in DSP processor and a microcontroller architecture for a significant cut in energy for the intelligence of installed machines. The SoC contains an analog to digital converter (ADC) sensor interface and extremely powerful PMIC circuits. The chip further consists of I2C, I2S, GPIOs, RTC, PWM, POR, SRAM, FLASH, and BOD. The patented hardware architecture (DIAL) is mixed with absolutely customizable CNN-based algorithms to carry out machine learning inference in heaps of microwatts.
The processor, named Tensai, can be used with the well-liked TensorFlow or Caffe Software. This solution can help various applications in audio, video and signal processing where energy is a strict constraint, including in UAV (unmanned aerial vehicles) sectors, within the Internet of things (IoT) and wearables.
ECM3531SP contains pre-trained learning machine speech recognition and keyword recognizing functions. ECM3531PG pre-trained photoplethysmogram (PPG) software, and ECM3531SF consists of machine algorithms for a fusion of gyro, magnetometer, and accelerometer sensors
The patented hardware architecture is mixed with the fully customizable Eta Compute algorithms based on CNN, GRU, LSTM, and SNN (spiking neural network) to carry out machine learning inference in only a few mW. Eta provides kernel software program for convolutional neural networks on Coolflux’s DSP, that is scalable in comparison with other NN (neural networks) and which can reduce a further 30% of the power with asynchronous know-how.