In an in-depth ipXchange discussion at CES 2024, Guy chats with GP Singh from Ambient Scientific, a provider of always-on AI chips with ultra-low power consumption. As GP explains, AI is a new type of computing that is different to the decision-tree-based systems that have evolved since the 1960s. If this is the case, why start AI chip development as a continuation of this lineage, especially when our brains do not work that way.
When finding a target piece of data, such as a user’s face, pattern recognition is key to (artificial) intelligence, rather than simple yes/no binary that asks whether each piece of data is the one the system is looking for. Like our brains, most AI chips work using pattern matching on every piece of data in the set, before discarding the irrelevant data. This wastes a lot of energy and computing time, so brain-like AI architectures are not necessarily the best way of implementing efficient AI. At least not in their current form.
Ambient Scientific uses a different approach to AI chip design by taking the computing evolution tree back to analogue circuit technology and reengineering the architecture from this point so that it performs AI tasks as efficiently as possible. The result of this startup’s efforts is an affordable, application-agnostic chip that uses single-circuit, multiple-analogue-matrix computing that still throws out the irrelevant data but does this with much higher energy efficiency than competing technologies.
For always-on AI on batteries, or even using harvested ambient energy, you can learn more about Ambient Scientific’s GPX-5 and GPX-10 AI processors by following the link to the board page. These highly integrated devices are a great fit for medical, wearables, industrial prescriptive maintenance, and in-cabin automotive applications, with up to 10x DigAn AI cores, 8 simultaneous sensor inputs, a Cortex-M4F core for non-AI workloads, and much more.
Keep designing!