Nice! Now I'd love to see the same design used for massive models. Current neural networks are pretty much a highly inefficient software emulation of the real thing.
AI chip adds artificial neurons to resistive RAM for use in wearables, drones
A newly published research paper describes a compute-in-memory (CIM) chip that combines artificial neurons with resistive RAM (RRAM) so that the AI model weights can be stored and processed on the same chip. A compute-in-memory chip based on resistive random-access memory A compute-in-memory chip based on resistive random- …
COMMENTS
-
Thursday 18th August 2022 13:09 GMT jmch
Local AI
If I would ever be using any "AI"* assistance, I would far rather that it resides in 'edge' devices, rather than send all my data to Amazon / Google etc for processing to get a result. It's far far better for privacy.
eg with Alexa, Siri and whatever Google's assistant is called, they are constantly listening in case someone says 'hey Alexa' (obligatoty xkcd**), and sending whoever knows what other data besides the audio stream for processing.
It might be much less efficient, get worse results, and hinder training of future "AI"*s but there's far too much snooping going on already!!
* definitely A but certainly not I
** https://xkcd.com/1807/
-