I wanted to know the max neuron configuration of AKD1000, so I...

I wanted to know the max neuron configuration of AKD1000, so I made a few calculations, scenario 1 with internal weights and scenario 2 with external weights (I assume that external weights are much slower as it will consume external memeory?):



I sort of came across some inconsistencies in the numbers, but there might be an explanation.

So, with internal weights the max configuration would be 18.4 million neurons (15 SoCs?), with external weights there could be 262.1 million neurons (yellow numbers in table) according to this source: https://brainchipinc.com/wp-content/uploads/2020/04/BrainChip_IP-Brochure_03302020_DD2.1.pdf

But if you connect 64 SoCs, you´ll supposedly have max 76.8 million neurons, which wouldn´t be possible with internal weighting if the above numbers are correct. So this must be with external weighting, but then there´s no way of getting to 262.1 million neurons, because of the max of 64 connected devices, which would leave you with a max of 76.8 million neurons: https://brainchipinc.com/akida-neural-processor-soc/

Further, what is the real number of Neurons per SoC? In the marketing literature it's 1.2 million. if it's with internal weights it's 360,000, because it has 80 NPUs (not nodes), if it´s with external weights it´s 5,120,000? On the other hand if we assume they made an error in the marketing literature, calling Nodes for NPU's, then there would be 1,440,000 neurons which is closer to the 1.2 million figure, but surely they wouldn´t have missed this glaring error?

Also, why would they advertise 1.2 million neurons on a SoC, a number that doesn´t fit with anything?

I can understand if there could be some limiting factors like addressing, but the numbers just seems so odd and out of place, couldn´t it have been optimized better, to enable better utilization of the resources?