Intel will be investing $13 million in Untether AI, a start-up building a cutting-edge microchip that transfers data a 1,000 times faster than a conventional AI chip. The highly conceptualized product is reserved for Artificial Intelligence and promises to perform neural-network calculations at lightning speeds.
Untether, which is based in Toronto, Canada, has already developed a prototype of this concept that transfers data between different components in a microchip a thousand times faster than a conventional AI chip. Although impressive, the prototype is far larger than the actual size of a chip and there are numerous factors that may impact the performance of the finished product.
One of the primary challenges with modern-day chips is moving data between memory to the units leveraged in performing logical calculations. The problem gets aggravated as the data scales up for AI chips to process capabilities such as face recognition. Untether is resolving this problem by integrating near-memory computing, which reduces the distance factor between the actual memory and the performing tasks. This arrangement helps in faster data transfers and lowers power consumption.
Untether is calling it the inference chip, which is different from the chip used to train large neural-networks in a data center. Instead, the start-up is keen to make a chip on a much smaller device, say a smartphone or a camera. Untether’s innovation may be a result of the recent boom in businesses leveraging Deep Learning abilities.
Intel’s investment may be a result of the company’s desire to penetrate in the ‘alternate chip design’ domain, an emerging technology being created to sync-in with smaller devices such as a mobile phone. Intel had, a few years back, acquired Nervana, another start-up developing chips for Deep Learning. Intel is on the verge of bringing the first products from this design in the market. The boom in AI seems to have caught on making corporations as large as Intel invest in them.