How woud an experienced person "know" how to build the circuit?

[T3sl4co1l]

Thats interesting and kind of how my mind works; to not think in the box but rather to define where the edges of the box are and see if problem lie inside or outside. Also good to see if a problem is too complex to figure out. When you say 4 chose 100 you mean there are four things that can each be connected up 100 ways? Would that be 400 or do they multipy together for some huge number?

They multiply... a lot:
https://en.wikipedia.org/wiki/Combination

"Choose" uses factorial, so it grows quickly.

Oh, also I got those backwards, it should read "100 choose 4", which actually makes sense, too.

So there are almost 4 million permutations, for an example circuit with an average* of 4 pins per net, and 100 pins.

*Actually, exactly 4.  If there are more or less on each net, while keeping the average constant, that gives more permutations again.

Actually actually... hmm, it's multiple choice, since there are multiple nets, after all.  Eventually you have either pins with no nets (or call them single-pin nets if you like), or pins in nets (with other pins), a partition problem.  Which works out similarly, but is a somewhat smaller (but still quite large) space again.

Needless to say, in-depth analysis through this route is quite useless; it's more just to say, if you tried it from the absolute bottom up, you'd have a hard time indeed.

Synthesis is indeed the way to go: to create something from the top down, going from, at the highest level, say, a rough specification and requirements, down to various levels of implementation, and finally the circuit itself.

Tim

[Beamin]

Thats interesting and kind of how my mind works; to not think in the box but rather to define where the edges of the box are and see if problem lie inside or outside. Also good to see if a problem is too complex to figure out. When you say 4 chose 100 you mean there are four things that can each be connected up 100 ways? Would that be 400 or do they multipy together for some huge number?

They multiply... a lot:
https://en.wikipedia.org/wiki/Combination

"Choose" uses factorial, so it grows quickly.

Oh, also I got those backwards, it should read "100 choose 4", which actually makes sense, too.

So there are almost 4 million permutations, for an example circuit with an average* of 4 pins per net, and 100 pins.

*Actually, exactly 4.  If there are more or less on each net, while keeping the average constant, that gives more permutations again.

Actually actually... hmm, it's multiple choice, since there are multiple nets, after all.  Eventually you have either pins with no nets (or call them single-pin nets if you like), or pins in nets (with other pins), a partition problem.  Which works out similarly, but is a somewhat smaller (but still quite large) space again.

Needless to say, in-depth analysis through this route is quite useless; it's more just to say, if you tried it from the absolute bottom up, you'd have a hard time indeed.

Synthesis is indeed the way to go: to create something from the top down, going from, at the highest level, say, a rough specification and requirements, down to various levels of implementation, and finally the circuit itself.

Tim

On a related note what is your IQ? I always am fascinated by someone who has a higher one then me, which is why I like this forum, very good to learn from. Less that one guy who thinks global warming is a hoax by the Chinese or something. 


Is this method how they map out neural nets to see how many combinations it can do knowing one is right and it has to go through all the combinations? What real world purpose do you use this for?

[TheUnnamedNewbie]

I always find it interesting to hear people talk about ICs as these magical blocks that are handed down from heavens. Someone also needs to design these ICs and things are somewhat different there. And it is not like there are just a few 'elite' designers working on these things - these are massive industries.

While they still use 'standard' topologies - to some extent a least - In my experience IC designers are much more familiar with the very low-level stuff, especially when talking about analog. Because of the limitations faced in that sector I think they go through a good bit more change in terms of technology and circuit design. Granted, I work in a millimeter-wave research group, and as a result the industrial partners I talk to are not exactly the ones that can make the same chip for 20 years at 5 cents a piece, but still.

Two things are different in IC design compared to circuit design using existing ICs as building blocks: 1) In these analog ICs you are working with those fundamental building blocks. You are not even designing with fixed transistors - you care actually deciding the dimensions (and perhaps type) of this transistor. As a result you need better understanding of how all these specifications relate. 2) You need to really understand what impacts performance because you need to fully simulate and verify the design.

With time you learn to really understand how to come up with new topologies and solutions. This is not something you can just learn in a university course, and takes years to master. I haven't really mastered it yet, though I have come up with some circuits of my own. As I work in a sector where we try to do what nobody has done before, we can't use existing building blocks because they just can't do it. That doesn't mean that our circuits are fundamentally different (though some are), but can mean we might spend more time designing parts that most designers take for granted (like studying the way to connect to the drain and gate of a transistor for 3 months because at 200 GHz, every little bit of resistance is a problem).

So how do we design a circuit? We start out with the concept - what do we want/need?
From there we do what everyone here has already said: Look at what building blocks we can use. Think about what specifications matter where: in the LNA we will care much more about linearity and noise than gain. In the PA noise is not a big issue because we are dealing with high signal power. In the mixer we might want to get high linearity and low LO leakage, etc.
We often even simulate using idealized models to get an idea of what specs we need for different parts. We can see what noise is acceptable in different sections, what linearity we need for the mixer, etc... and how it all influences each other.

From there on we look at how we can design the blocks. We try to meet the specs for every block, but if this is not possible, we need to look at why we can't (over time designers will also get a bit of 'instinct' about what specs are realistic. We can start from an understanding of different topologies, we know what topology will be good at what. EG, for a variable gain amplifier I can look at making an actual variable gain amplifier. However, this might be harder to verify performance-wise, and could be more challenging to control. I can use a fixed gain amplifier with variable attenutators in between, which could be easier but requires more power. Perhaps multiple amplifiers in parallel, and switch on the amount we need to give us the right gain? But this will take a lot of area because we need to duplicate this amplifier again and again.

[T3sl4co1l]

On a related note what is your IQ? I always am fascinated by someone who has a higher one then me, which is why I like this forum, very good to learn from. Less that one guy who thinks global warming is a hoax by the Chinese or something. 


Is this method how they map out neural nets to see how many combinations it can do knowing one is right and it has to go through all the combinations? What real world purpose do you use this for?

Neural nets are very general, as is the method of analysis applied to circuits.

Namely, rather than trying to simplify the circuit (not too hard a problem, as the permutation space can be reduced quickly -- I guess that would count as a projection of sorts, into a smaller space), you just write down every possible connection, in a matrix, and solve that.  The problem reduces to linear algebra, which while tedious, computers are exceptionally good at.

Projections are interesting: they are a mathematically consistent way to... avoid doing more math.  The result of a projection has fewer dimensions than its input.  You've thrown away data, but also (potentially?) made it more manageable.  Projecting a 3D space onto a 2D screen is a classic example.  Perspective tricks -- ambiguity -- is precisely the consequence of throwing away data.  You can't tell how big and far away something is, just that it takes up some apparent visual size.  You can try to infer from surroundings and such (as our eyes are so good at doing), but there's always some edge case that can fool it (an Escher drawing for example).

Neural nets are heavy on linear algebra.  The map of connections is also a matrix: a grid of numbers, each column showing the gains from a given node, and each row showing the gains to a given node. When a network is "trained", all that's happening is, each of these numbers is tweaked up or down incrementally, to give a better output.  "Better" being a matter of comparing against reference data -- a given input-output pair, an example of the problem, paired with its solution.  Neural networks are very dumb indeed, they have to be shown millions of examples before they give accurate results.

Tim

[tggzzz]

So how do we design a circuit? We start out with the concept - what do we want/need?
From there we do what everyone here has already said: Look at what building blocks we can use.

I agree with everything you have said, but there is also an alternative to that particular point.

Sometimes, particularly in research rather than product development, you:

• see a new component/technique and ask how it can reduce problems with existing components/techniques. A classic example is high speed high resolution ADCs enabling digital SDR receivers
• see an existing component/technique and ask how it can help in a completely different application domain

Both of those require familiarity with existing practice and imagination - and there's no short cut to getting those.

[chris_leyson]

So how do we design a circuit? We start out with the concept - what do we want/need?
From there we do what everyone here has already said: Look at what building blocks we can use

Exactly that ^^^
Over time you build up a mental tool box of circuits and/or techniques and as an engineer that's what you carry around in the grey matter. Once you've defined the problem you can sketch out a few circuit blocks or if it's unfamiliar territory you can search for a solution and see how other engineers have tackled the same problem. You may very well have a lot of other design parameters to think about, for example, performance, cost, size, efficiency, reliability etc. It all depends on the initial concept, sometimes it's just a case of putting together standard building blocks and sometimes you have to think "maybe I can combine some circuit techniqes and thereby build a better moustrap", but will it work ? If you can model and simulate your design then that's a bonus but sometimes there are things you just can't accurately model without spending a lot of time doing research and in that case it's quicker to build it and test it.

The other day I needed to add a monostable multivibrator to a design but it has to be teeny tiny, an LTC6993 in a 6-pin SOT23 package would fit but it needs programming resistors as well, costs about $3. Use an Atmel 6-pin ATtiny and you've got 2 monostables for $0.3 maybe, prices are ballpark. Sometimes thinking outside the box pays.

Beamin, just build stuff, poke it with a scope probe and literally see how it works or doesn't. You will learnt a lot more from things that don't work because you have to investigate why they didn't work and that way you can get a better insight or maybe intuitive grasp of how things should work. Gut feeling and intuition are also very important design tools but they have to be learnt, sometimes the hard way.

@Beamin, thanks for posting the Widlar Hassler video, I had completely forgotten about that design, maybe you can do it with an electret mic, an op-amp or transistor for a bit of gain and 6-pin or 8-pin micro with an A-D, just another way of solving a problem but in that application you get a lot more options by going digital, maybe some crude spectral analysis, just to target that really annoying colleague.