We were initially expecting Akida 2 to include LSTM, and this would have required the inclusion of additional memory in the NPUs.
Then more recently we were alerted to transformers, which are alleged to be able to understand natural language. To do this, they will need sufficient memory to capture the context of speech, requiring a much larger memory. This means that each die (SoC) will have a significantly larger silicon footprint on the silicon wafer.
We were told that the electrical circuit design of Akida 2 with LSTM had been sent to Anil's group quite some time ago, before transformers raised their head(s). So it is possible that the extended period between the finalization of LSTM and this announcement was the result of the later addition of transformer functionality (and concomitant additional memory and associated logic).
So I'm if, as a result of the larger footprint, wondering if Akida 2 will be produced in a smaller format, eg, 7 nm, which, as the table shows, as of 2023, is quite a mature technology, so mature that a conservative company like Renesas has chosen it for its Akida MCU?
After all 7 nm has been with us for over 4 years (
https://en.wikipedia.org/wiki/7_nm_process#:~:text=The first mainstream 7 nm mobile processor intended,was released at Apple's September 2018 event. ).
View attachment 26652
There were sound reasons for choosing 28 nm for Akida 1, and I'm surmising that the immaturity of smaller scales was among the foremost. However Renesas' choice of 7 nm suggests that the tech is now well accepted.
Of course there are a number of intermediate technologies which Anil may choose, but I doubt he will go smaller than 7 nm.
fixed
