Yesterday I attended a developer conference in South Florida. One thing I noticed is that several of the presentations this year involved AI, either AI in art, education, or writing production applications. One presentation I found interesting was on spatial computing and AI.
The presentation was done by a member of the HoloLens team, who described the components, technology, and practical use of the device. The FDA has approved it for surgical use, and they are being designed for the military --both training and field operations for HUDs. The see-through screens on the device are the most expensive to produce due to their manufacturing failure rate. It was pointed out, however, that devices that use passive video pass-through would not be considered for these types of uses, because if they were to lose power, the user would be in the dark.
Current components in the device mentioned were CPUs, and GPUs, while newer hardware like TPUs and NPUs were being researched. TPU - Tensor Processing Unit; NPU - "Neuro Processing Unit".
I asked a question of the presenter, specifically, "Given the concerns for privacy and security, and that most AI inferencing for small devices is done in the Cloud, are there plans on improving the hardware to run AI locally on the device, especially since it may be used where connectivity is not available."
The answer to that research was being done in this area, and NPUs offered high-speed processing while consuming less power. The current devices already face challenges with battery life and thermal considerations. After the talk, I asked her if she was referring to neuromorphic computing, and she said "Yeah, that's it." I didn't attempt to pry, but I mentioned a few of the current processors I was aware of, ending with Akida and some of its impressive specifications compared to other neuromorphic hardware. As she is a hardware developer on the team, it never hurts to put the bug in one's ear.
This tells me that companies know they need this technology and are aware of it. Research into neuromorphic computing hasn't stalled, it is taking some time. While BrainChip has IP available now, getting it from IP to chip form will be a long process for customers who have not started the process.
As the
SoC market continues to grow, I think BrainChip has made the wise choice of making the IP available. Companies can order SoCs with several features à la carte tuned to their specific device needs to allow better control over costs, necessary functionality, and energy efficiency.
Hopefully, BrainChip can maintain the talent pool needed to keep innovating Akida and stay ahead of the competition until neuromorphic technology gets a foothold in the consumer market.