Hacker News Reader

Comment by westbywest | Hacker News Reader

Analog electronic computers actually offered significant advantages in precision over their digital equivalents for much of the mid-20th century, due to limitations imposed by slow clock speeds and small word sizes. Calibrating them was non-trivial, though. https://en.wikipedia.org/wiki/Analog_computer#Modern_era

View on HN

In case anyone enjoys reading about the history of control systems, Between Human and Machine: Feedback, Control, and Computing before Cybernetics by David Mindell is so much fun to read.

https://www.amazon.com/Between-Human-Machine-Cybernetics-Tec...

Interesting, especially in light of the "modern" view where the prevailing narrative is that analog computers are (in general) less precise, but faster. And, of course, less flexible, but I think that goes without saying.

I am curious to what extent an analog/digital hybrid can blend the best of both worlds, especially for something where the speed <-> precision tradeoff may somewhat favor speed. Like training neural networks.

Everything is analog when you look at a small enough time scale. "Digital" is an abstraction on top of analog, not a substitute.

It appears that digital is king for modern calculations. However, digital exists only in the ~millivolt to 5V range, its difficult to get a purely digital circuit to "sense" microvolts or nanovolts.

That's where Analog comes in. You can shape and calculate microvolt and even nano-volt level signals before upscaling them to millivolts so that the CPU can handle the rest of things.

In general, digital RAM will hold a value more reliably for a longer-period of time than an analog storage device (aka: capacitor). True, RAM is implemented using capacitors (DRAM needs to be "Refreshed" as those capacitors lose value).

Its easy to "refresh" a capacitor if you only care about two states: high and low. If you're analog however and care about microvolts / nanovolts, its impossible to "refresh" capacitors.

After all, what does a capacitor at 5000uV really mean? In digital, you can refresh it back to 6000uV every few seconds. In analog, you don't know if part of the information leaked out or not.

At the boundary between analog and digital, you typically need to provide an anti-aliasing filter.

If you are measuring a differential signal with a large high-frequency common-mode component, then it also makes sense to perform the subtraction in analog. Otherwise the inevitable small phase shift between the sampling of the two channels in your ADC will end up coupling the common-mode signal into your difference.

While what you say is true, its incredibly amazing how good ADCs are in the modern era.

3 Gigasamples / second: http://www.ti.com/lit/ds/symlink/adc07d1520.pdf

The fastest OpAmp I can find is this 18GHz GBP: http://www.ti.com/lit/ds/symlink/ths4303.pdf

At 3 GHz, the OpAmp would only offer a gain of 6 which isn't really enough for much accuracy.

Granted, the OpAmp is like $5 and the ADC is hundreds of dollars (and the ADC is only 7-bit accurate)... but the digital world has gotten scary fast and scary good.

That's why Oscilloscopes, even GHz-Oscilloscopes are being made with digital technology today.

Hell, this crazy product raises eyebrows: http://www.digikey.com/product-detail/en/analog-devices-inc/...

26Giga-samples (Nyquist of 13). I mean, 3-bits sucks, but holy crap is that fast. A 26GHz ADC would be able to perform digital-filter analysis on 2.4GHz Bluetooth and Wifi without any aliasing what-so-ever.

Like anything, its a cost tradeoff. DSP can be much cheaper than active analog for many frequency ranges. You can certainly oversample your way to a cheaper antialiasing filter in many applications. But you can't ever eliminate it entirely.

That exists. It's called mixed-signal ASIC's. They mix digital and analog components.

Reply on news.ycombinator.com