Comments by "Maggie Jetson" (@maggiejetson7904) on "The Most Powerful Computers You've Never Heard Of" video.

Analog is good for things that need no precision (i.e. AI) but cannot run with too many transistors or too much power from too many clock cycles.
6
@norbert.kiszka As an electrical engineer I have to say, the calibration process will be a bitch if every transistor must be calibrated to do analog computing right.
2
@norbert.kiszka I know how opamp works, but you know they are +-5% unless you calibrate them right? How many do you think you can stack them together without going way off? Not thousands of them for sure. If you want precision resistors they too need a lot of calibration. Your cheap transistors are also +-1% at best or 5% at worst without calibration.
2
@norbert.kiszka Silicon variation from the fab can vary +- 5% (Qualcomm MDM modem chip for example, but 5% is very typical as manufacturer's guarantee), unless you bin the heck out of them and that means calibration as I mentioned. Are you just a student in school with no real world experience? Just because you get a few thousand chips that's consistent does not mean it is guaranteed by the vendor, they can change all of a sudden if they are not guaranteed.
2
@norbert.kiszka "if You want do digital calculations of analog data (temperature, weight, whatever) then You need analog circuits - if You like it or not." That just means you need an analog circuit that will be calibrated when convert to digital number before you do any calculation. Your converter will be calibrated and your whole system will be calibrated before you send it out of the factory. Now if you are trying to do an all analog computer (no A2D then do all digital calculation), then each step needs to be calibrated or else your weight of your "op amp" inside the chip will have whatever error % to the power of whatever you stage you decided to stack together with. To compete with digital calculation you will likely need at least a few hundred if not thousand stage, and even 0.1% to the power of 1000 could easily mean huge magnitude difference. How are you going to deploy one system to another when all the weight are different? Oh resister is expensive on a silicon, so you can bet the better way to do the amplification (calculation) is to adjust the voltage to the gate of the transistor directly (which means calibration too). Have fun, when temperature change these calibration needs to be accounted for too.
2
Shannon is the most influential guy in our world in the last 40 years, seriously. So many equations and theories from him.
1