Re: Rewriting SSA. Is This A Chance For GNU/Linux?

On 06/04/2025 05:26, c186282 wrote:

On 4/5/25 9:07 PM, The Natural Philosopher wrote:

     temperature conditions that add errors in.

>
Not really, That was mostly sorted years ago.

     Ummm ... I'm gonna kinda have to disagree.
    There are several factors that lead to errors in
   analog electronics - simple temperature being
   the worst.
 

Not really. If you se as close to zero coefficient resistors and enough feedback the circuits  are insensitive to temperature.
In terms of using junctions to do multiplication, there are ways of compensating for temperature in one device using the same effect in another, and on chip they are thermally coupled and doped the same. So it works very well indeed

 

   Keep
   carrying those errors through several stages and soon
   all you have is error, pretending to be The Solution.

>
So no different from floating point based current climate models, then...

     Digital FP *can* be done to almost arbitrary precision.
   If you're running, say, a climate or 'dark energy' model
   then you use a LOT of precision.
 

And get a very accurate 'wrong answer'.
The problem is the time it takes to do it.

 

   Again, perhaps some meta-material that's NOT sensitive
   to what typically throws-off analog electronics MIGHT
   be made.
>
   I'm trying to visualize what it would take to make
   an all-analog version of, say, a payroll spreadsheet :-)
>

An awful lot of op-amps.

     To say the least :-)
    CAN be done, but is it WORTH it ???
    But, I suppose, a whole-budget CAN be viewed
   as an analog equation IF you try hard enough.
 

Most complex dynamic systems are 'analogue' and anyone who has modelled them using analogue electronics can tell you that if they do not have the right negative feedback they become unstable, and that's how we engineers know that 'positive feedback' in the climate is bullshit.
What you can do with a multitude of opamps resistors and capacitors is to model extremely complex dynamic systems very quickly, and look for instability or unexpected results.
No one does it any more, because everyoine fell in love with digital, but there are times when it works better.

 

The thing is that analogue computers were useful for system analysis years before digital stuff came along. You could examine a dynamic system and see if it was stable or not.

    Well, *how* stable it is ........
    Digital is always right-on.
 

No, it isn't.

   So what do you NEED most - speed or accuracy ?
 

If not you did it another way. People who dribble on about 'climate tipping points'have no clue really as to how real life complex analogue systems work.

    I'm just gonna say that "climate" is beyond ANY kind
   of models - analog OR digital. TOO many butterflies.
 

That is the point.
There are chaotic elements but its *bounded* chaos.  I looked everywhere for mathematical papers to try and determine the criteria for a chaotic system to become bounded, or even to identify how many strange attractors there were, or where they were located.
No one has done the work.
And yet digital climate modelling cannot even represent the past, let alone the future.

   Now discrete use of analog as, as you suggested, doing
   multiplication/division/logs initiated and read by
   digital ... ?
>

Its being thought about.

    And we shall see ... advantage, or not ?
 

In certain cases it is a better solution. I can envisage a chip comprised of many many linear amplifiers whose gain, frequency response and interconnections were programmable by digital logic, to allow one to model an enormously interconnected system very quickly, to at least see what its sensitive areas in fact were....
I.e. identify the butterflies...

   Maybe, horrors, "depends" .....
 

Well that means nothing. That's just anti-tech speak for 'I don't know what I am talking about, and I am as good as you. so therefore you don't either'...

   The "real world" acts as a very complex analog
   equation - until you get down to quantum levels.
   HOW the hell to best DEAL with that ???
 

The point is you don't.  If your system is so unstable that one atomic decay renders the cat dead, it doesn't last long in the 'real world
The real world is *conservative*. It consists of systems that have *temporal persistence*. They have 'stood the test of time', if you like.
If your model - be it analogue or digital - doesn't have the right sort of feedback to do that, its clearly not an accurate model is it?

   Oh well, we're out in sci-fi land with most of this ...
   may as well talk about using giant evil brains in
   jars as computers  :-)
>

Well no, we are not.
Digital traded speed for precision.

     I'd say digital traded precision for speed ...
 

No. Digital is SLOW. Many hundreds of cycles to do in what analogue can do *approximately* in one or two.

Massive parallelisation will definitely do *some* things faster.

    Agreed ... but not EVERYTHING.
    Sometimes there's just no substitute for clock
   speed and high-speed mem access.
 

Well tough, because you aint gonna get that . The speed of light is the speed of light

Think 4096 core GPU processors...I think that's the way it will happen, decline of the general purpose CPU and emergence of specific chips tailored to specific tasks. Its already happening to an extend with on chip everything....

    I kinda understand. However that whole chip chain
   will likely need to be fully, by design, integrated.
   This is NOT so easy with multiple manufacturers.
 

Oh of course.
".

    ANYway ... final observation ... it keeps looking
   like we're far closer to the END of increasing
   computer power than the beginning. WHAT we want
   computed is kinda a dynamic equation, but OVERALL
   we're kinda near The End.
    THEN what ?

I think we have taken in around 75 years, Turing's basic concept of a Turing machine to the end of the line, more or less. There is still room for improvement, but not at the level of general purpose computing cores.
The lowest fruit left to pluck are the massive parallelisation  which will suit some tasks - like matrix multiplications where the same job is being done to enormous pieces of data and the answer of one does not affect the result in another core - that benefit.
As primarily an engineer who cut his teeth on analogue systems and built analogue computer systems at university, I am extremely interested in the potential use of hybrid chipsets for modelling of all sorts of real word complex dynamics systems, from the economy to the climate, from faster race cars with more downforce, to better aircraft and sailboats.
What all these systems are is a massively interconnected series of 'amplifiers' with varying gain response, and phase delay, whose transfer function may indeed not be linear.
We know the partial differentials - e.g. Navier Stokes equations - but that doesn't help in digital terms. It's simply too complex to model the interactions accurately at any scale. But a chip with 50 million opamps and multipliers on it, all individually controllable that could be linked together in different ways could really start to do the job digits cannot.
In short a new sort of computer for the problems that digital computers simply do not work on.
Think about it. Division is slow in digital. But in analogue a logarithmic transfer function, followed by a subtraction and an exponential, does a rough division in nanoseconds
Any mathematical transfer function takes the same amount of time, because instead of calculating it, it is precompiled and built in to the silicon Better still it may be possible to describe a non linear transfer function digitally and make the amplifiers in the chip obey it. Using the inherent non linearity of a transistor operated near its cutoff. Yes you curve might well be a series of 256 short straight lines or so, but it will still be FAST.
No one has done this. They have stuck with the very mathematical approach of Turing and the applied maths of summing of series etc, for which a digital computer is admirable.
What I am describing however is the equivalent of a book of log tables - remember those - baked in silicon, and being applied to opamps as transfer functions
Why it isn't already done in terms of floating point maths on digital computers I do not know. A simple look up table for a log value, adding and subtracting exponents, plus an exponential table and you have a two or three operation multiplier or divider. Sine tan and Cosine tables likewise...Its no good for operation on integer matrices, but its a bloody fast way of getting a reasonable accurate engineering answer  - like using a slide rule, really.
My point is that real world analogue problems - i,e, not counting how many cents the bank has in its deposit box - but calculating at what point a car tyre is going to start to slide, and what happens next, even if subsequent runs give a different answer due to noise, at least tells you that the result OF that tyre sliding is unpredictable and chaotic, so try not to do it...is still giving a very valuable piece of information.
The day that a computer beats a wind tunnel, is the day I look forward to. That is a possible way to do it
--
I would rather have questions that cannot be answered...
...than to have answers that cannot be questioned
Richard Feynman