Well, I finally got around to watching the video (see link below) that Pom posted a few days ago with Jonathan Tapson.
It’s highly technical and, to be honest, a fair bit of it went over my head. That said, there were a few comments that struck me as potential red flags, unless I’ve completely misunderstood what he was saying.
He starts by outlining the core features of neuromorphic chips such as asynchronous, event-based architectures and so on (see slide below). But then he makes the point that asynchronous designs account for less than 1% of commercial electronics, adding that it’s probably closer to 0.00001%!!!!
I could be wrong, but hat doesn't even seem niche to me, it seems effectively non-existent in mainstream commercial terms.
He also notes that just because the performance advantages are clear doesn’t mean adoption automatically follows. In other words, that technical merit doesn’t equal market penetration.
On top of that, Jonathan mentions how very few machine learning engineers have experience with spiking neural networks. So even if the hardware is differentiated, the developer ecosystem is extremely small. Add to that companies’ reluctance to share proprietary datasets and internal models, which he acknowledges makes integration difficult and you start to see how the barriers to adoption stack up.
He does however, mention BrainChip has found a way around the challenge of customers not wanting to provide their proprietary info to third parties, but he emphasises that it's hard work. He also mentioned something about being happy at some earlier decisions BrainChip made, which should help.
Unless I've got this all wrong, I think this video might help explain why commercial traction has been so much slower than many of us hoped.
I’d welcome input from someone with stronger technical expertise, in case I’ve completely misinterpreted any of this. The last thing I’d want to do is draw the wrong conclusions since we're all depressed enough as it is.
Here are some random slides from the preso.
