Digital design is hard. But in the right environment, digital circuits are more forgiving than analog. That 3.3V signal coming out of the chip has to drop a lot along the way to not be a logic level at the destination. If you are trying to push the boundary then digital design has much of analog design, but mostly you get a bit of a pass on many things that plague analog designers. Berkeley’s AI research group has been experimenting with using deep learning to evolve analog IC design.

Analog ICs are plagued with noise sources and often don’t have the margins that digital circuit designers enjoy. According to the post by [Kourosh Hakhamaneshi], designers often build a few blocks and attempt to lay them out in a way that should work and meet other requirements. Then they employ simulation, make changes as required, and simulate again. Accurate simulations can be very time intensive. You can read the actual paper, too, should you want to dig into the details.

Continue reading “Darwin Approves: Berkeley Evolves Analog Design” →