Intel Invests In An AI Chip That Can Transfer Data At 2.5 Petabits Per Second
AI chip startup, Untether AI (founded in 2018), came out of stealth on April 1st and received $13 million in Series A funding from Intel Capital and other investors. The startup, which has delivered over 1 billion chips to the market, is developing a new chip that eliminates the data movement bottleneck that costs energy and slows down performance in traditional architectures.
Untether AI has already developed a successful prototype that shows that there is a possibility of higher efficiencies and better performance for AI accelerators. The commercial product that is being developed now is coming at the right time as the AI inference market is projected to grow to $22 billion by 2022.
The startup’s Kensington architecture gets data to the processors at 2.5 petabits per second, which is 1000 times faster than what a traditional architecture can do. Untether AI uses near-memory computing which aims to reduce the physical distance between the memory and the processing tasks. One of the reasons this is so fast is because the chip eliminates the need of using a bus to transfer data, which can improve efficiency. This also means that there is an opportunity to use the chip in a larger scale with less energy being required to run it.
Traditional architectures followed something that is known as Moore’s law, which is the observation that the number of transistors in a dense integrated circuit doubles about every two years. This happens because companies just scale down the size of their transistors on their microchips. The smallest transistor created is 1 nano metre long, so they are getting very small and could soon reach a limit where it becomes very difficult to shrink the transistors anymore. This means that new ways to improve the speeds of microchips are being developed, such as Untether AI’s new chip.