2021 BRN Discussion, page-27306

Thanks for sharing Ph1960oz, interesting article. I agree with FF, the last sentence was quite interesting:



I think it's the first time we've received any details on what Renesas are doing with their license.

A few things worth considering here:
-They are only licensing 2 nodes (8 NPU's). This means it isn't intended for high end applications. If you look back to the Rob Telson presentation at the Edge Auto Summit, he tells you how many NPU's were used for each application in the driver classification scene (see screenshot below).
Here it used:
-14 NPU's for person detection
-50 NPU's for face identification
-7 NPU's for keyword spotting.
This indicates that they are licensing Akida for a low end board. So it is probably intended for applications where latency doesn't matter but where power consumption does (Akida can run larger networks on fewer NPU's, it will just run slower). Thinking for example about an application where latency doesn't matter, the fridge rotten food smell detector comes to mind. If a sensor is in the fridge it can take all the time it needs to detect rotten food. Usually food rots at a slow process and people don't need to know the instant food begins to rot, as it occurs over a lengthy process. That said, I'm sure there are many other applications where this MCU could be used where latency doesn't matter.

-If you think back to the licensing fee in the previous 4C (~400k or 500k from memory), this has been confirmed as being for a 2 node system. Obviously this relies on factors like volume, but this indicates it is roughly the minimum amount Brainchip would expect for a license. For a single-use license at that. If customers want to license more NPU's or for multi-purpose applications, they will likely have to pay more.

Pure speculation, DYOR

Edge Auto Summit demonstration (someone enters a car, it identifies someone is there, recognised their face and does KWS).