Semiconductor startup Mythic has launched an artificial intelligence (AI) processor that it says will deliver the computing capabilities of a GPU for a tenth of the power.
The M1076 Mythic AMP chip targets edge computing AI applications with power and footprint limitations. Like other edge-based AI processors, the Mythic CPU runs a neural network to handle a process called inferencing. This takes new input, such as a video signal from a camera in a manufacturing plant, and runs it through a trained AI model designed to recognize specific things — like an object on a production line.
The difference lies in how the chip processes data in its neural network. It uses a combination of analog circuitry and in-memory computing to save power. Rather than storing its data in slower memory that the chip must access when needed, it uses multiple memory arrays that handle the computing themselves using their memory elements as tunable resistors.
This allows the chip to compute neural network inputs and outputs as voltages and currents. The memory arrays function as nodes in the neural network graph and can operate in parallel.
The unique chip architecture enables the company to fit more AI computing horsepower into a single chip, it says. The M1076 delivers up to 25 trillion operations per second (TOPS) using a 3-watt power draw. It's available as a single chip, a PCIe M2 card for low-footprint applications, and a PCIe card with up to 16 chips. The latter can deliver 400 TOPS for just 75 watts of power.
The AI-powered supply chain
Better demand forecasting and operational excellence
FREE DOWNLOAD
To run neural networks trained by popular AI packages such as PyTorch, developers will need to retrain them to run on the chip's Analog Compute Engine (ACE) and then compile them for the chip's architecture.
Mythic says the chip will work well in video processing applications, including object detection and depth estimation, making it suitable for industrial machine vision, surveillance cameras, and augmented reality applications.
The company raised $70 million in series C funding in May, led by BlackRock and Hewlett Packard Pathfinder, bringing its total funding to $165.2 million.
