2021 BRN Discussion, page-5967

Hi MC,

Well spotted.

Glad to see you spotted my deliberate mistake - just shows you can't believe everything you read.

I was drawing my information from slide 4 of the presentation I posted in my first reply to you.

https://d1io3yog0oux5.cloudfront.net/_e6768070d76a9b211503e7500d571de0/brainchipinc/db/236/1703/pdf/BrainChip+Linley+Akida+Presentation_v5.pdf Slide 4

which was dated 2019. I guess that we have now found there are some instances when converting CNN to SNN where we need 8-bit weights to maintain sufficient accuracy. Changing from 4-bits to 8 bits is equivalent to the difference between slicing something into 15 pieces and slicing it into 255 pieces.

We still use 4-bit activations.

I imagine there would probably be a trade off of speed and power for the increased accuracy when using 8-bit weights.

Changing from 4-bit weights to 8-bit weights would have required a significant redraft of the Akida layout and would have contributed to the 12 month delay.

ARM is still 20th century von Neumann computing.