Maxim introduces neural network accelerator chips to enable IoT artificial intelligence in battery-powered devices

     Maxim Integrated Products, Inc (NASDAQ: MXIM announced the launch of the MAX78000 low-power microcontroller with a neural network accelerator that enables battery-powered embedded Internet of Things (IoT) devices to make complex decisions at the edge through fast, low-power artificial intelligence (AI) reasoning. Compared to software solutions, this fast, low-power decision implementation reduces the energy consumption of complex AI reasoning to less than one percent of previous solutions, and battery-powered systems using AI technology can significantly extend their uptime, helping to enable a new generation of battery-powered AI applications that were previously uninsurable. In addition, the MAX78000 does not affect latency metrics and costs: it costs a fraction of FPGA or GPU schemes, while performing inference 100 times faster than software schemes implemented on low-power microcontrollers.

     AI technology enables machines to see, hear, and perceive the world in ways that were previously completely impossible. In the past, laying out AI inference to the edge meant collecting data from sensors, cameras, and microphones, then sending the data to the cloud to implement inference algorithms, and then sending the results back to the edge. This architecture is challenging for edge adoption due to high latency and power consumption. As an alternative, low-power microcontrollers can be used to perform simple neural network operations, but latency is compromised and simple tasks can only be performed at the edge.

     By integrating a dedicated neural network accelerator, the MAX78000 overcomes these limitations, enabling machines to see and hear complex patterns by performing AI processing locally and in real time at low power consumption. Because the MAX78000 performs inference with less than one percent of the power that microcontroller software runs on, it significantly improves productivity for applications such as machine vision, voice, and facial recognition. At the heart of the MAX78000 is dedicated hardware designed to minimize the power consumption and latency of convolutional neural networks (CNNS). The hardware runs almost without any microcontroller core intervention, which means that the operation is streamlined. Energy and time are used only to implement the math of the CNN. To efficiently feed the collected data from the outside world into the CNN engine, users can use one of two integrated microcontroller cores: the ultra-low power ARM Cortex-M4 core, or the lower power RISC-V core.

     Given the challenging nature of AI development, Maxim Integrated provides tools for a smooth evaluation and development experience. The MAX78000EVKIT# includes audio and camera input, an out-of-the-box presentation platform that supports large-word table keyword retrieval and facial recognition. Extensive documentation helps engineers train the MAX78000 network using the tools it uses every day: TensorFlow or PyTorch.