Linley Falls 2020, page-91

In response to Blue Ant posting the link:

Thanks BA for providing the link. That was the most interesting presentation about Akida so far, imo. The further, deeper explanation about how the mesh network NPUs work together with the breakdown about activation sparseness and weight quantisation reducing the number of MACs required to run CSNN was amazing to me. I think this technology is so far ahead - we are out there on our own with viable edge use case solutions for just about anything. What stuck in my mind was Anil's description about the mesh and how communication between the different layers of the network is done entirely in that hardware - whereas conventional approaches require a CPU and separate memory to save values attained between layers. Man when it is explained in that way you can appreciate this is really game changing and there is currently nothing else like it.

One thing though which I haven't seen others comment on, Anil states that the Akida chip samples only have 20 nodes? Now forever the mesh network is described in the Akida architecture as specified with 64 nodes, but he says 20 in the presentation? The demo examples of hand gesturing (using MajikEye), ENose and the ball bearing example don't require many nodes - certainly not 20 nodes of computation. So maybe they've figured for the samples or even the first production batch that 64 nodes is not necessary? Maybe his description of nodes differs from NPUs and where I'm thinking 64 is still valid? Dunno, anyone else got a theory on this?

jtard.