As I pen these words, I recall once constructing a speech synthesizer as part of some project I was toying with. The digital portion of the design was relatively straightforward and everything appeared to be functioning as planned. When it came time to perform its sole function in life, however, my pride and joy sounded like a frog with severe head cold (as opposed to the rich, dulcet tones I was expecting).

It didn't take long to resolve the problem to the analog portion of the design (a smattering of components intended to filter, condition, and amplify the output from the circuit), but at that point I ran into a mental wall. Not having any clue how to proceed, I asked one of my analog designer friends to take a peek. After a few minutes glancing at my circuit diagram and muttering incantations to the analog gods, he rooted around in my "box of bits," pulled out a capacitor, touched its terminals to what appeared to me to be totally unrelated portions of the circuit...and voila...it worked!

To this day, I still don't have a clue as to what he did or why he did it. I'm sure that I learned something valuable, although truth-to-tell I'm not exactly sure what it was. I guess that the main lesson I took away with me was that analog design was best left to someone else (a simple rule of thumb, but it's served me well for more than 20 years).

The problem is that the world doesn't remain still. The number of mixed-signal designs (those containing substantial amounts of both analog and digital circuitry) are rising at an exponential rate. Furthermore, although the analog content of today's SoC designs typically accounts for less than 5% of the total transistors, these transistors reflect around 20% of design aspects such as silicon real estate and power consumption, and are associated with around 50% of design re-spins.

The problem with analog
The real problem is that analog designs contain too many variables, and their solution spaces are too complex, for most of us to wrap our brains around. For example, consider a black box consisting of two knobs that control a light. When the indicator lines on each of the knobs are pointing straight up, the light is at its brightest, and when either of the knobs are rotated to the left or right the light starts to dim down.

The solution space for this problem is very simple. If we were to visualize this solution space as a three dimensional diagram, it would look like a small curved hill. Even if you didn't know the rules, if someone asked you to work out the knob positions resulting in the brightest light, it would only take a little trial and error to determine the optimal solution.

But now consider what happens if we have say 50 knobs and 50 switches, where all of these devices are interrelated. For example, turning one knob clockwise may brighten or dim the light depending on the state of a number of the switches. Even worse, each of these switches may individually have different effects on other knobs, the position of which may affect the actions of other switches, and round and round the circle we go.

This is very similar to an analog design, in which tweaking the size of one transistor may affect the actions of another, which modifies the results from an inter-related portion of the circuit...and so we plunge deeper and deeper into the mire.

In this case, the solution space starts to look like an alien landscape. If we are lucky, it may resemble a rugged terrain with lots of mountain peaks and valleys. The problem being that even if we detect a local peak, how do we know that it is the biggest (corresponding to an optimal solution). Even worse, the solution space may contain weird constructs like "bridges" or "cities in the sky," whose names are reasonably self-explanatory.

There are algorithms available that can be used to attack this sort of analog design problem. For example, "simulated annealing," which is based on the way in which liquids freeze or metals re-crystallize in the process of annealing. A simple way to visualize this is that you seed everything with random values (corresponding to a high temperature where everything is disordered), then make a number increasingly smaller iterative changes (corresponding to things cooling down) until you reach a steady state.

Then you "throw everything up into the air" and start again with new seed values. You do this for as long as your patience (and computing power) holds out, and then take the best solution in the desperate hope that it actually corresponds to an optimal circuit configuration.

A much more exciting alternative is to use so-called evolutionary (genetic) algorithms. The idea here is that you start of with a seed "population" of solutions, which you evaluate and rank. Then you discard most of the poor solutions and you "breed" the more successful solutions by combining portions of different solutions in different ways. You also throw in some level of random mutation, evaluate and rank the new population, "breed" a new generation of solutions, and off you go again.

Surprisingly enough, this evolutionary algorithm-based approach is incredibly efficient at evaluating extremely complex solution spaces, such as those associated with analog designs. Of course, like most things this sounds easy if you say it quickly, but actually putting it into practice requires people with size-16 wide-fitting brains, and where are we to find guys and gals like these?

Enter ADA (stage left, fanfare of trumpets, roll of drums)
Well, by some strange quirk of fate, a few years back the folks at Analog Design Automation Inc.(ADA) were pondering how to apply their vast knowledge of such esoteric topics as evolutionary algorithms, neural networks, and fuzzy logic. One problem that was crying for attention was the complexity of analog designs, so they leapt into the fray with gusto and abandon.

A few days ago, ADA launched their Genius V2 family, which comprises two core products: Creative Genius V2 and IP Explorer V2. I was lucky enough to see a real-time online demonstration of these little rapscallions, and it was enough to make my eyes water. Armed with these tools, even a digital drongo like myself stands a chance of creating optimal analog designs.

First of all we have Creative Genius V2. This commences by reading a circuit topology, testbench(s), and process models from your existing design environment (e.g. Cadence ADE or Mentor DA-IC). You then use Creative Genius V2 to specify which parameters and attributes you can change and how to change them (including min and max values).

One really cool thing about this tool is the ease with which you can specify your target goals -- these are things that you can measure like open loop gain, phase margin, power consumption, and so on (remember that analog designs typically have a lot of goals, which often conflict with each other).

Creative Genius V2 then uses its evolutionary algorithms to tweak all of the parameters associated with the analog circuit, evaluate the results, and breed and converge on a range of potential solutions (some of which will be better in certain areas and worse in others). Furthermore, you can use Creative Genius V2 to evaluate a selection of alternative topologies.

The key point here is that its use of evolutionary algorithms means that Creative Genius V2 is a high-capacity optimizer than can explore a richer solution space with 5 to 20 times more variables than traditional low and medium-capacity optimization engines, which are based on linear-gradient and simulated annealing algorithms, respectively.

Multi-dimensional visualization
Of course Creative Genius V2 suffers from the same problem as many other EDA tools, in that it generates vast amounts of data that is almost impossible to wade through without some form of computer-aided visualization. This is where IP Explorer V2 leaps into the fray.

Figure 1 -- The IP Explorer V2 interface

Each line of text in the upper-left window corresponds to a particular circuit solution generated by Creative Genius V2. The big window at the bottom contains what ADA refer to as a "Parallel Axis Plot." Each vertical bar in this plot represents some "goal" such as phase margin or open loop gain (green is good, gray is less than the minimum spec, and red is critically bad). Each solution from Creative Genius V2 is represented by a multi-segment line linking the various goals.

In the upper right hand corner is a more traditional X/Y co-ordinate display window that allows you to plot "this" against "that" (where "this" and "that" are any two variables or goals). You can quickly and easily use the "this versus that" window to filter out solutions in the other two windows by simply dragging cut-off boundary lines around the window.

There are so many cool features to IP Explorer V2 that I don't know where to start. For example, you can have as many "this versus that" windows as you like, each plotting different variable pairs, and you can "tear off" these windows and leave them running wild and free on your screen. Similarly, you can do the same thing with Parallel Axis Plot windows, and one trick I really liked was the ability to normalize all of the results in a Parallel Axis Plot to the initial "seed-value" circuit to see where and how different solutions compared to the original implementation.

As usual there is way too much to cover in this short column, but if you are involved in this type of design you would be well-advised to bounce over to ADA's website to download more detailed information, datasheets, and suchlike. In the meantime, both Creative Genius V2 and IP Explorer V2 receive an official "Cool Beans" from me. Until next time, have a good one!

Clive (Max) Maxfield is president of Techbites Interactive, a marketing consultancy firm specializing in high-tech. Author of Bebop to the Boolean Boogie (An Unconventional Guide to Electronics) and co-author of EDA: Where Electronics Begins, Max was once referred to as a "semiconductor design expert" by someone famous who wasn't prompted, coerced, or remunerated in any way.