Ambient Scientific frees AI from the cloud with analogue in-memory compute architecture

What it means for hiring

The number to test is the 70 to 75 per cent. If that much of the energy in an edge inference goes into moving data rather than computing on it, no process node fixes it and the win has to be architectural. Analogue in-memory compute attacks exactly that, and it has been promised before, so the question is whether this one ships inside a customer's product rather than a demo board.

Either way the hiring problem is the same and it is severe. Analogue and mixed-signal designers who can also read a neural network graph barely exist, because the industry spent twenty years telling analogue people their skill was legacy. The firms doing this well hire senior analogue designers out of data converter and sensor front-end work and teach them the machine learning. That direction works; teaching a machine learning engineer to design analogue silicon does not.

Read the original at eeNews Europe .