Imec and KU Leuven show chip with digital and analog coprocessor for AI tasks

Researchers from imec and KU Leuven have designed a chip that combines a digital and analog coprocessor. The analog accelerator is more efficient than the digital one, but the digital accelerator is needed for computational precision and programmability.

The imec research center has mainly developed the chip for artificial intelligence tasks in, for example, augmented reality glasses, autonomous drones and smart robots. These devices often have to process a lot of environmental images, which according to imec are best processed on the device itself for privacy reasons. However, this also requires a lot of computing power, which means that the batteries of such mobile devices can quickly run out.

For this reason, Imec developed an analog chip, AnIA, two years earlier, in which calculations are performed directly in the memory. As a result, ‘most of the calculations are performed ten to a hundred times more energy-efficiently’ than by a digital accelerator.

However, AnIA was less suitable for tasks where computational precision and programmability were important. That is why imec and KU Leuven have developed the Digital and ANalog Accelerator, or Diana. It combines a digital and analog accelerator and chooses which coprocessor can perform it best based on the specific task. For example, pattern recognition is best performed on the analog accelerator, while processing those observations is better done on the digital coprocessor.

According to the researchers, Diana combines ‘the best of both worlds on one compact chip’. GlobalFoundries manufactured the chip. It is not clear whether the chip will be used in practice. The researchers want to present the results of their research at the Isscc electronics conference next week.