The Comdyna GP-6 analogue computer that I use daily at work | Hacker News Reader

The Comdyna GP-6 analogue computer that I use daily at work

reddit.com

64 pointspmarin37 comments

This thread had a very good OP. I don't buy that this is at all useful though. It would be trivial in a modern computer to get much more precision than this. I feel like his arguments about conversion from analog to binary would have merit, if the scale that modern computers operated at wasn't so severe. Yeah conversion from analog to binary can be lossy, but the losses become smaller as you use more memory. And these days we have access to a lot of memory.

What's interesting is that our modern computational platforms have capabilities that far exceed most of our use cases. And this is a good demonstration of why.

From what I've gathered and read from OP's comments, it sounds like their analog model makes sense and is precise enough for them to work with. People talk about how much more powerful a "32-core Xeon machine" is and they're not wrong, but just because new technology is available you also have to look at is it worth it to transition over/update it.

I have an example from my daily work. We were hired to update a company's model from 6 years ago. Their model is an Excel VBA model that we're rebuilding in Python with an increase resolution of detail and accuracy. However, for our "more accurate" model to be accepted, we have to also prove the following:

- We have to show that our model's outputs are the same as the VBA's model for the same input/starting condition (makes sense at a higher level, but at a higher resolution this means you have to spend more time looking at why a specific value is in a specific way and can be different from the lower resolution model)

- Train the new users on the new model (knowledge transfer)

- Ability to update and modify the new model wherever they need to change it or understand it

- Understand the limitations and the constraints of the new model

This basically boils down to how much time is needed in training the new model, how much money and time is needed to invest in building a new model, and is it really worth it to build a new model when the old model has worked good enough for now (and you understand the limitations of said model and is within your needs)? There's only so much time in the day and you obviously have a ton of other tasks that also needs your attention, and getting these newer models verified and accepted by the board is great but it's also additional meetings and time that you need to setup with them and get everything OK-ed... doubly hard when everyone else is also super busy.

I'd say modern computational platforms are better when it comes to raw power, but the logic and just resources required to build, verify, accept, and train (human resources-wise) the new model might just not be fully worth it.

Remember, there are also other industries other than tech that uses computer hardware to get jobs done.

I agree that updating tech stack can be expensive and sometimes not worth it. Being on the bleeding edge isn't a great idea for a business use case. You have to draw the line somewhere. I do think that the line is drawn very well ahead of analog computers though.

I think that this guy is using the analog computer because he's an enthusiast. That's fine. It's not better by any means other than making an employee happy.

The thing about conversion from analog to binary is that it's a one-time-cost only. With analog computation, you pay for every single computation/conversion step with increased noise, even if each operation theoretically provides more precision than the binary equivalent. There's just no equivalent to the noise-resiliency of binary compute.

On the contrary, every single digital floating point operation has the inherent capacity to add noise. To make things worse, the noise depends very much on order of operations, relative magnitudes of operands, and other algorithmic aspects of what is usual an estimation method for solving equations.

Estimating accumulated error in floating-point calculations is an extremely difficult science. On the other hand, a simple circuit that solves an ODE in an analog computer can be analyzed rather easily.

So don't use floating points? You can use arbitrary precision math libraries.

If you devote enough memory, you can be much more precise with continuous values than an analog computer could ever hope to be. Frankly it wouldn't take much memory at all by modern standards though.

Arbitrary precision doesn't help with irrational numbers.

You may be thinking of the "just use enough bits so you're sure the accumulated error isn't significant" approach, which is essentially what modern floating-point attempts to do.

There are still some pitfalls though, around doing an analysis that proves the bounds of said error in a given algorithm and data set.

You might be interested in the concepts used in interval arithmetic [1] and specifically the work of Gustafson in designing some models for using those in scientific computing hardware.

[1] https://en.wikipedia.org/wiki/Interval_arithmetic

Analog computers aren’t all that unusual. For example, many airplane pilots use a Jepperson Circular computer to calculate headings and other parameters.

https://www.amazon.com/Jeppesen-Circular-Computer-diameter-J...

Even Spock used it on the Enterprise. https://www.flickr.com/photos/10901121@N06/1196208122

Well, a slide rule (even a circular one) is not an "analog computer"...

How about the E6B's wind vector calculating elements? Those pieces are on the reverse side of the slide rule bits.

https://en.wikipedia.org/wiki/E6B

Still no, an analog computer is an electronic computer which uses varying voltages and currents to calculate results. It's not the same as a manual computer.

Analog computers need not be electronic. They can be mechanical (e.g. https://en.wikipedia.org/wiki/Digi-Comp_I) or hydraulic (e.g. https://en.wikipedia.org/wiki/MONIAC), too.

(Generic entry point: https://en.wikipedia.org/wiki/Analog_computer)

I appreciate the OP (of the reddit post) wants to keep his/her identity secret but I remain completely unsatisfied with the response to someone asking exactly what they need this analogue computer for

> My job is to make and maintain models of economic trends in my company’s industry. I find analogue computing to be the best way for me to easily and effectively model dynamic real world systems.

I don't have much of an understanding of analgoue computers so I can't grasp why this computer is better at solving this sort of problem over a digital one.

> so I can't grasp why this computer is better at solving this sort of problem over a digital one

As the original author mentioned in the discussion, it's better for _them_ and that's enough in that situation to keep doing it that way. The assessment may be entirely different for you, or maybe even everybody but that author.

It hasn't been established that it's better, only that he or she thinks it's better--and we're all trying to understand why that is. Because we're (or at least I am) 99.99% sure that it's much, much worse.

To start with, the whole thing could easily be simulated in software...

I suspect that the main reason why it is better is UX of representing whatever you want to calculate by physical structure and tunable parameters by physical knobs.

I don't have a good answer, but there does appear to be a history of using analog computing for financial modeling.

One famous example is the MONIAC, which used fluids to simulate the flow of money: https://en.wikipedia.org/wiki/MONIAC

I wonder if 'fuzzy' or uncertain models like economic ones can tolerate or benefit from slightly less deterministic algorithms.

My favorite “impossibly difficult” computing problem that’s unsolved (brute force required) by digital but solved by analog is in finding the shortest path from A to B, wherein you just apply electricity to the start and endpoint and let the neon glow demonstrate the optimal solution. For example:

https://www.nature.com/news/2002/020520/full/news020520-12.h...

While this is certainly a great trick (my favorite version of it is the use of soap bubbles to solve the traveling salesman problem) this solution is certainly no better than what you would achieve with an optimization that runs on a digital computer (and very often it is much worse, given the advances in numerical optimization).

There is one very general problem with these "analog" solutions to optimization problems: nature (i.e. the analog computer) quickly finds a local minimum and it does not bother to find a global one (or as the sibling mentioned, it finds a stationary solution, not the true minimal solution).

There are tricks you can use to try to search for a better minimum, but all of these tricks are applicable to the digital computer as well. Which is really exciting: you get to talk about annealing and how it works the same in metallurgy and in numerical optimization. Even more fascinating, you get to talk about how a whole host of diseases are inherently due to this "failure of convergence" of analog computers: prion diseases would not exist if analog computers (our cells) folded their proteins to the true minima.

I'm pretty sure the actually difficult (i.e.: not handled by the heuristics TSP solvers use in practice) examples for TSP are impractical to convert to a physical object, and I'm not sure that physics would actually do a good job with them. Remember, it's the principal of _stationary_ action, not _least_ action, so it's basically a greedy search.

Finding the shortest path from A to B is very much a solved problem, e.g. Dijkstra's algorithm can do it in quadratic time on arbitrary graphs.

The article tries hard to pretend that the analog method presented could solve traveling salesman problems, but the actual paper only talks about ordinary shortest path problems. I don't see how it could be extended to TSP either, since it's hard to make electrons actually visit all destinations you want to cover when there's a shorter route available.

Apparently the Navy also used mechanical 'computers' on old battleships to calculate firing solutions.

Sailors would adjust cranks and shafts to set the input values, and those inputs were connected to cams which were machined in the shape of solutions to the equations that needed to be solved. You can watch an old training video about it from the '50s:

https://youtu.be/s1i-dnAH9Y4

Bombers too.

https://en.m.wikipedia.org/wiki/Norden_bombsight

> I wonder if 'fuzzy' or uncertain models like economic ones can tolerate or benefit from slightly less deterministic algorithms.

It depends on what sort of fuzziness or uncertainty you're talking about. If you're referring to a model's sensitivity to initial conditions / measurement noise, then a common solution is to use Monte Carlo simulation, in which case you get an estimate of a probability distribution function as an output (which is often collapsed to an expected value, but that's not necessary). I'm not aware of any analog computers that give probability distribution functions as outputs rather than single values.

Maybe the OP works in infosec, and this is the one computer they can guarantee won't be hacked.

I would assume this computer generates a huge amount of EM interference(?) that could be mined with an antenna up to a mile away.

The frequencies involved are generally low enough that EMI is not an issue (which is generally required because of big nests of random banana plug patch cords of unknown RF performance)

To me OP comes off as a curmudgeon that won't learn how to do this more efficiently in software. There's no reason why they couldn't do it faster and more accurately in software.

I kinda feel the same, but I'd probably put a bit more of a charitable spin on it. I can certainly see how it might be more intuitive to the OP to use an analog computer and how it might not be worth the cost of moving to a software system when this is good enough.

What I would disagree with is any assertion that this is objectively the best way, and OP does acknowledge in the original thread that this isn't necessarily the overall optimal setup.

That's a fascinating read.

Analog computing may make a comeback:

https://spectrum.ieee.org/computing/hardware/not-your-father...

As some of the commenters on this reddit thread mentioned, any electronic musician who has a Eurorack rig containing, for instance, a Make Noise Maths, technically uses an analog computer regularly at work.

How fascinating. I wonder what is the measure of precision on such computers for real numbers. And can this be used for integer math if at all ?

Closest that I've come is working with opamp circuits. Generally speaking: Doing almost anything in analog is more expensive than doing it digitally. These days, the only things you do in analog are things that are impossible to do digitally, like anti-aliasing filters. RF is another story, but even there the economics are changing with direct-sampling ADCs.

1% resistors are easy to find. 0.1% can be found at a high price. 0.01% can also be found, for roughly 1 USD/ea, and only on the E12 sequence. Capacitors are worse by a factor of 5 or so. Those tolerances turn into scale-factor error, common-mode rejection ratio, DC bias, and so on. Without breaking the bank, you can easily build a circuit to maybe 7 bits of accuracy, and for a small timer you can do maybe 12 bits or so. But in volume, its almost always better economically to use a lower-accuracy circuit and perform calibration at the factory or in the field.

Aside: I fact-checked myself on Digikey and was pleasantly surprised to find that you can get 0.01% resistors without "calling for quote." A few years ago the limit for small-timers was effectively only about 0.1%. Even still 1 USD/ea is a BOM budget-buster for all but the most narrow of uses.

Circuit with an ENOB > 12 definitely exist. See table 1 in the app note below for an example. https://www.maximintegrated.com/en/app-notes/index.mvp/id/53...

ADC's well above ENOB exist. 24-bit sigma-delta ADC's are pretty easy to come by without breaking the budget.

But analog computation circuits (subtraction, gain, addition) are another matter entirely.

Doesn’t one make a sigma delta from subtracting, integration and gain stages? If one could make a sigma delta one could make other stuff too.

https://en.m.wikipedia.org/wiki/Delta-sigma_modulation#/medi...

Sigma-delta ADC's are built in such a way that component errors result in changes in bandwidth instead of changes in scale factor.

Reply on news.ycombinator.com