Brain-like hardware accelerator provides power-efficient AI inferencing by ignoring irrelevant data
In ipXchange’s next AI-focussed interview from CES 2024, Guy chats with Todd about BrainChip’s Essential AI computation architecture.
Where many ‘AI’ chips are simply a matrix multiplication engine – essentially just doing quick maths rather than truly intelligent thought – BrainChip’s solution attempts to replicate the way the human brain operates. By recognising important data – from a sensor, for example – and ignoring irrelevant data, BrainChip’s solution operates with much less memory and less computing power for higher efficiency than mathematical solutions that process the complete dataset.
Todd then shows us an example application with a Time-of-Flight (ToF) sensor performing distance and depth measurements to recognise users in a way that is much harder to trick than a 2D AI imaging system. In this example, BrainChip’s solution takes the data from the sensor – i.e. Nolan and Guy’s faces – and classifies them after quickly training the model to recognise the subtle differences of their features. This is done with respect to their shoulders for even greater certainty of user identity. This type of setup might be used in industrial and in-cabin automotive applications for user verification, safety – such as for user-specific airbag deployment – and security.
BrainChip’s solution works as a neuromorphic – i.e. brain-like – AI hardware accelerator in conjunction with a host processor. BrainChip also provides the software and IP solutions to put this architecture into fully integrated chipsets, but with development systems preloaded with standard demo use cases, design engineers can easily get started using this technology for low-power edge inferencing.
Learn about BrainChip’s Akida AKD1000 SoC by following the link to our board page, where you can apply to evaluate this technology for use in a commercial project.