TI degrading specs of existing parts without changing part numbers

In the original National parts, all four channels would go bonkers if any input went slightly below ground.

And the test sets.

A support guy said "That's an old Burr-Brown part. We don't know much about those."

Speaking of TI, this is incredible:

John Larkin Highland Tech Glen Canyon Design Center Lunatic Fringe Electronics

And saturated the tail source transistor, which then hogged all the base drive from the other sections.

They fixed that in the 324A and the 358, iirc. Doesn’t seem to happen anymore, at least.

Nice. We’ll have to find a use for it.

Cheers

Phil Hobbs

A frustrating thing is that the opamp manufacturer has an exceedingly accurate spice model of the part that their designers created in order to optimise and verify the design, but the company won't give that one to the public. The schematic in the datasheet is often drawn by marketing people who have never discussed it with the design engineers, let alone seen the real schematic, and is sometimes a source of great amusement to the actual designers. The public spice model is quite possibly created in a similar manner, but possibly corrected based on some measurements by apps engineers or based on complaints from customers. Meanwhile the really accurate model sits on some internal server being ignored while the design engineers have moved on to some other product.

Don't bother.

Whatever that is supposed to mean.

If the simulation is any good. The published simulation models are mostly behavioral, and precise enough for the purpose the marketing people had in mind.

It's "commercial in confidence" because it exposes a lot of the detailed design of the chip being simulated.

Commercial secrecy pretty much demands that.

The public spice model is quite

It's not ignored - just too revealing to be made public.

The late Jim Thompson used to claim that he could create behavioral models that were just as accurate as the transistor level models and a lot more compact. He was touting for business, but he had to be able to deliver something pretty good to make it worth his while to advertise at all.

The ears are close to the brain for a reason. Golden ears don't seem to be close to particularly good brains.

Why would they bother? You can create the old patterns with much higher accuracy on 8 - inch wafers with modern tool, and many of the old designs worked fine if merely shrunk a bit (creating even more parts per wafer).

There are several commercial lab services that will take a competitor's chip that you obtained, (including mixed-signal designs much more complicated than an op-amp), and reverse engineer it for a smallish fee. They provide detailed schematics and layout. It is much cheaper than a mask set and well within the means of every competitor. It isn't usually that interesting except perhaps e.g. to see whether one's patent is being infringed.

Keeping the best spice models secret from customers achieves little, as any competitor would already have full schematics if they wanted them. Perhaps the non-technical management and shareholders don't appreciate this, or don't appreciate the significant extra value to the customer of having a simulation model that is really accurate.

The LT Spice library part models seem to be pretty good. They are behavioral models, and I assume that they have internal, semiconductor-level models, that run really slow.

I found one LTC part that, with no power connected, would generate a teravolt on one of its pins. Free energy! Obviously they use ideal current sources in the public model.

John Larkin Highland Tech Glen Canyon Design Center Lunatic Fringe Electronics

A smallish fee isn't going to be all that small.

If they wanted them badly enough. Reverse engineering may not be as expensive as a complete mask set, but it is quite expensive, and keeps fairly skilled people from your own firm busy for quite a while.

I worked o voltage contrast for a bit, and the kind of electron microscope that can see what's going on on the exposed surface of an ude-encapsulated chip isn't cheap.

And really slow? Behavioral model are not only less revealing than transistor level models, but run a lot faster.

"Understanding" seems to be missing here.

Probably not true. I suspect that it is Laplace transforms all the way down.

Crude mathematical models can generate large numbers if appropriately stimulated. Rationalising this as "ideal current sources" is imagining an unnecessary extra mechanism.

I read a study that says that bats can correlate time between their ears with nanosecond resolution. Pretty good for wet stuff.

John Larkin Highland Tech Glen Canyon Design Center Lunatic Fringe Electronics

If they don't own the fab, there may well be NDA restrictions on the device models.

Cheers

Phil Hobbs

Improbably good. I knew a psycho-acoustician who studied bats

Good point. For most of the processes I used, we made the models, including some processes at external foundries. I expect this might no longer be so common, especially for foundry CMOS processes.

The person who arranged it said it was a few thousand. A mask set is maybe a hundred times that.

No, just pay someone to do it. But, it's not all that useful if you intend to make something much better than what your competitors already make.

I think at least some of those reverse engineering companies are run out of university labs. They already have the equipment and it's not always busy. The service I know of did not operate the chip, they just took images and made netlists / schematics, and measured the process parameters.

It is good to have both. My block on one chip simulated 1e12 times slower than real time, but that had many thousands of transistors in it, and computers have got faster over the 20 years since then. I did have to simulate the whole real circuit for one beat period of a two-tone distortion test, and running all the process corners and temperatures did take many days, but I also made a fast behavioural model to make sure that the digital state machine could calibrate the analogue part properly etc.

It's never just "paying somebody to do it". You've got to spend time in house making sense of the results. And making something much better than your competitors already usually depends on knowing why they are as good as they are

Modern chips have several layers of metalisation. "Taking images" can involve taking off a layer or two of metalisation to expose what's underneath. University labs aren't always as well equipped or as well-informed as they like to think. They have just as many pretentious wankers as every other organisation - perhaps more since their stock in trade is presenting stuff at conferences.

Circuit designers don't really care what's going on inside the chip. A behavioral model is usually perfectly adequate.

Much of the progress has come from putting lots of processors in parallel.

That's process validation rather than circuit design.

Bats are pretty good, but that seems a bit too good. Can you find that study?

Thanks,

Joe

Good grief, just google

bat echolocation nanosecond

The history of science is rich with people saying "that's not possible" when it turns out to be.

John Larkin Highland Tech Glen Canyon Design Center Lunatic Fringe Electronics

Google finds claims of as low as 10ns, but I have not yet found an original reference for such small time intervals. On the other hand, the abstract of this paper claims

Perception of Echo Phase Information in Bat Sonar James A. Simmons Science 22 Jun 1979 Vol 204, Issue 4399 pp. 1336-1338 Echolocating bats (Eptesicus fuscus) can detect changes as small as 500 nanoseconds in the arrival time of sonar echoes when these changes appear as jitter or alternations in arrival time from one echo to the next.

John

And hear a lot of unsupported opinion. Like all that nonsense on circuit design.

I studied bat sonar quite intently at least 20 years ago, and saw that bat sonar was very good at finding moths on the wing. Some bats used CW (good for detecting moths hiding in vegetation) and Chirp FM (good for moths out in the open), anticipating human designed radar and sonar by a hundred million years or so. But as others have said this did not require nanoseconds, and I don't recall such numbers.

It did use correlation, but in a neural map that was spread out on a

I have a thick folder on this somewhere.

Oh absolutely. Biology will find a way to achieve dinner, at least on average.

Joe