I promised somebody to get back on the answer regarding apparent...

I promised somebody to get back on the answer regarding apparent discrepancies in the numbers of neurons on Akida, I only got the answer that their support staff would get back to me shortly... a month ago. Do any of the patent-divers here come across something that can shed some light on the discrepancies?

So there are 80 NPU´s on their SOC:
https://brainchipinc.com/akida-neural-processor-soc/

The NPU´s have either 4,500 or 64,000 neurons each, depending on whether they use on-chip memory or external memory.
https://brainchipinc.com/wp-content/uploads/2020/04/BrainChip_IP-Brochure_03302020_DD2.1.pdf

So either 360,000 neurons or 5,120,000 neurons.

or

1,200.000 neurons?
https://brainchipinc.com/akida-neural-processor-soc/

Do they expect that the applications use a mix of on-chip and external memory?

If Akida is to operate independently without external memory, then the real number of neurons is 360.000 I suppose?

Also the memory requirement of the neuron should also depend on whether it´s 1-, 2- or 4-bit weights, else it wouldn´t make much sense to have the option of 1- or 2-bit weights? So, if the NPU only supports up to 4,500 on-chip weighted neurons, this would likely be at the lowest weight of 1 bit? So, a 4-bit configuration would get closer to 90.000 neurons?

If for whatever reason the 5,120,000 neuron figure (with external weights) is only at 1-bit weight and somehow a 4-bit weight would lower the number of neurons to about a quarter (could be an addressing limitation), then we get to 1,280,000 neurons which is fairly close to the 1,200,000 million specification, but that´s pure speculation.