Electronic design

Jan 21, 2024 Last reply: 2 years ago 13 Replies

JL wrote an interesting post in the depths of the "better microelectronics from coal" thread that I thought was worth pulling out on its own.



On 2024-01-21 10:12, John Larkin wrote:>



"...what IS electronic > design, and what's the best way to do it? <snip>




> Short answer, cobbling. When presented with a problem or an > opportunity to design electronics, the most efficient way to do that > is to grab a piece of paper and immediately sketch a circuit or an > assembly. Sometimes one can do that instantly, without thinking, or > sometimes one can ignore the issue for a few days and then the design > pops up. Sometimes brainstorming and whiteboarding help. Sometimes > fiddling with Spice helps. >
> All that literature research and math analysis and simulation and > breadboarding and prototyping are just slow and expensive follow-up > chores for people who don't have 100% confidence in their instincts. > Analysis, sometimes prudent to do, but not design. >
> Design is subconsious and instinctive. And it's free! And to some > extent, it can be taught, but seldom is. >
> Most of us design things to sell, so do whatever works. We're selling > stuff, not publishing papers. >

Hmm. I don't think that I agree in general, because you make it sound as though the process were just intuitively plucking one idea out of somewhere-or-other and cranking it out.


You've often argued in favor of brainstorming, where you get a few smart people in front of a white board and try out ideas to find the best one and flesh it out. We've done that together, very fruitfully.


It's possible to do more or less the same thing by oneself, but it requires the ability to tolerate uncertainty for extended periods. (That's a skill well worth developing, which most people are really, really bad at, IME.)


I sometimes need to do a family of designs, rather than just one. Recently I've been working on some very fast, very cheap SPAD preamps, intended to go in the guts of positron-emission scanners.


Designs with lots of real-world constraints are often the most fun, and this one's specs include: 300-ps edges with 100-ps timing repeatability from unit to unit; no magnetics allowed; and a BOM cost of $1 or less. (You need a whole lot of channels, and PET and MRI machines are often combined.)


I do a fair amount of analysis of circuits of that sort, to figure out what actually limits their performance. It isn't super detailed--in this case, just enough to figure out whether it'll be the base-emitter time constant, the Miller effect, or the SPAD's series resistance that will be the limiting factor.


Miller, I can deal with using circuit hacks. The BE time constant is Rbb' * Cbe, which gets slightly worse at high current, but is mainly a device parameter--to get a big improvement you have to change transistors. The SPAD can be negotiable depending on whose process you're making them on--when each machine needs thousands of them, vendors tend to listen.


Eventually, of course, you have to pick one and go with it, but picking a topology usually takes me an iteration or two.


Cheers


Phil Hobbs

It's the same idea that your FNG suggested in the design review two weeks ago; the one that you shot down in flames.

RL

I agree with both of you. What Phil is doing is figuring out where to focus the brainstorming and fiddling, and the resulting wild alternatives can easily be assessed. It's at the very least an orthogonal method.

My personal experience is that iterations and inspirations require studying extensively followed by sleeping on it, so the metric isn't a few days, it's a few nights.

The bit about the necessity of nights was pointed out by J. Hadamard in his famous book on this issue. The book has become hard to find and expensive, but has now been reissued:

.

formatting link

Joe Gwinn

So, you're all wet?

Actually, I also get ideas in the shower, probably because I stopped focusing so hard.

I used to keep a waterproof dictation recorder handy, and on my bedside table, so I wouldn't lose the ideas, but don't need the recorder any more.

But the key is to stop trying for a while and think irrelevant things.

Joe Gwinn

I don't think so. Just a few words are not enough to specify and generate a specific, reliable design.

There's not much I in AI. It's mostly a silly fad.

But brainstorming isn't so-called. Done right, it really works.

It helps if one has a rubber duckie (or maybe I'm just that bad at it!)

That's not intelligent. It's just automating a lot of grunt work, as programmed. Line monte carlo simulation. The person who set it up is the intelligence. All that's new is having more compute power than we had in the past.

Computers automate grunt work and let us work faster and better and move up the abstraction stack. Nonlinear differential equations were never much fun.

I would like a Spice that was, say, 500x as fast as mine is now, nvidia or something. And I'd love some way to specify results and have a program juggle values and even library parts for a best solution.

Past attempts at such optimizations have tended to diverge. Even most interns are smarter than that.

Take a look at flux. It's funny.

Depends on what you mean by “ideal.“ For a pure LTI system, I agree. However, even SPICE isn’t really LTI—besides roundoff and truncation error, it’s implemented using floating point, which has magnitude limits.

You can easily use a nice noiseless behavioral amp whose gain is a weak function of the time-averaged amplitude.

That’s an idealized model of the HP200-style ALC.

Cheers

Phil Hobbs (Off to a second day of deposition in a patent case.)

John, AI is NOT just more computing power.

It is neural networks running on hardware that work well with NN

I see that with the Agent filter you put out that stops the Google Groups postings stops the Fred Bloggs and other responders evidently. But I haven't seen any sporge posts either so that is good.

boB

[snip]

Yes and no. No computer hardware "works well with NN" (Neural Networks), because nobody has invented a true direct associative hardware memory yet, so the computer hardware roughly emulates the desired NN kind and behaviors, at stunning expense in hardware and electrical power (plus cooling systems to remove all that heat).

It is useful to note that the pacing task for implementing a neural net in a digital computer is matrix inversion, where the matrix to be inverted may be billions of lines by billions of rows, and is not at all sparse. It's a long story, but well documented.

By contrast, the human brain has a volume of about 1.3 liters and consumes about 20 watts, and contains something like 171 billion cells, of which 86 billion are neurons. Computations are performed by analog hardware, cells. The number of synapses per neuron is something like a factor of ten thousand larger. Every synapse needs a cell in the matrix holding at least an 8-bit value. And so on.

Joe Gwinn

If it runs on a compuer it sure is.

Do any NNs work well? NNs are cargo-cult crude cartoons of an actual organism. Single- and few-cell organisms without a nervous system learn and do complex stuff, and our brain has billions.

And a human brain can play tennis, or recognize one face out of a million, or design circuits, with wet chemistry gates that have millisecond prop delays.

Different kind of computer.

Well, in this case, using multi-core graphics controllers is organized way better for the die size allowed in the neural network application. Its different than what you can get these days in a particular die size for the particula application. A PC micro, although as or more complicated than an Nvidea graphics processor, it works better for that.

As for being THE processor for a NN like Joe better points out, these processors are what we had (and have) at the time so they were used.

Just wait for a just as complicated, circuitry wise, neural network chip somes out. It will be even better.

Anyway, NNs are different too in that they "learn" better than a "program" running on a 100X faster PC chip. That's what I considered being different than just more computing power. In a non-linear way mabe.

boB

"In London, where Southampton Row passes Russell Square, across from the British Museum in Bloomsbury, Leo Szilard waited irritably one gray Depression morning for the stoplight to change. A trace of rain had fallen during the night; Tuesday, September 12, 1933, dawned cool, humid and dull. Drizzling rain would begin again in early afternoon. When Szilard told the story later he never mentioned his destination that morning. He may have had none; he often walked to think. In any case another destination intervened. The stoplight changed to green. Szilard stepped off the curb. As he crossed the street time cracked open before him and he saw a way to the future, death into the world and all our woe, the shape of things to come."

formatting link

I opened an email in the morning and took a shower and had a bunch of ideas. So ideas both get delivered in the suds, and happen there too.

Other people have noted the creative powers of hot water falling on your head.

Good book, First Steps by Jeremy DeSilva. It's about the evolution of upright walking, but he mentions that various great thinkers had ideas sleeping, showering, or walking. Walking works best in the woods, not on city streets.

Join the Discussion

Have something to add? Share your thoughts — no account required.

Didn't find your answer?

Ask the community — no account required