Miniaturization

Yeah, that was a black eye for LT, back it the day when datasheets were eng ineering documents and not marketing ones. My faves in that class are the A DA4898 and 4899.

Cheers

Phil Hobbs

Oh, and now everybody does that trick, at least with chopamps and things like single-supply dpots, some of which have charge pumps inside.

Cheers

Phil Hobbs

Examples, please. Unless the "industrial equipment" chip is a custom ASIC, this would be very unusual.

For pretty much everything, today, right down to SOT-23s, the cost of the package is a large part of the cost. Add in test and it's easily half the cost.

Your total costs will go way up. No one is going to want to package

Very interesting - but I don't see myself building a bandgap reference from a few dozen 3046 packages heating themselves to equal temperature.

I have used transistors on 3046 arrays to heat them and to measure the temperature, but never the same transistor. (Beforehand I measured the transient response of the thermal coupling between all possible pairs, to see what loop behaviour to expect. They are different, as you would expect, as the transistors are not all mutually equidistant from each other.)

I got balled up on microns and mils... back in those days it was mils... spacing was 0.1 _mils_ minimum, or 2.5um

This discussion has deviated from hobbyist into quantity. Restore an old K&S Wirebonder and do it yourself.

I did that around 1977. Bought two at a junkyard in San Jose and made up one working unit.

With CoB, why do you need a custom chip?... roll your own system with off-the-shelf chips. ...Jim Thompson

Neither would I. But you don't have such good resistors on-chip. Current mirrors (especially large-ratio ones) are horribly noisy anyway. (A 1:1 mirror produces an output current noise 3 dB over full shot noise, and that 48:1 monster would be 20 dB above shot noise.

The basic problem with bandgaps is that you have to put > 20dB gain on the PTAT part for it to cancel out the V_BE drift. I'm much happier with buried zeners.

In a built-up circuit, you have a lot more bypassing available, so it's really only the low-frequency noise of the reference that you care about.

In some ways, board-level and chip design are like real space and Fourier space--different things are easy and hard.

The MAT14 is a symmetrical quad, so it works better. Fancier versions of the laser noise canceller use one diagonal pair for the difference amp, and the other two as heater and temperature sensor. That makes the gradient effect nearly zero.

Cheers

Phil Hobbs

"Jim Thompson" wrote in message news: snipped-for-privacy@4ax.com...

Let's see, wire bonders are on eBay from $500 (to $15k ;-) ). As long as they *work* at that price, that's not unreasonable, and the per-part cost should be small.

As for chips, suppose I need something smaller. Suppose I want to beat Phil at his own game by shamelessly copying his laser noise canceller circuit onto a chip. When performance of a discrete circuit is dominated by parasitics of the board (this may not be the best example, but others apply), but you don't sell a million of them, what do you do?

Tim

All you need is a customer ;-) ...Jim Thompson

Hybrids can help, but the parasitics want to scale like length, rather than area, so shrinking the circuit by a factor of, say, 10 in area only gets you a factor of ~3 better parasitics. (Not that that isn't a lot.)

Wire bonding is a skill I've never mastered myself. IIUC you need to screw up several dozen times before you get passably good at it.

Cheers

Phil Hobbs

I used it in NMR gradient coil drivers where, out of sheer luck, the noise bump didn't result in significant magnetic field. It's a nice opamp, low voltage noise, low drift, very fast. Current noise is awful, but I was amplifying a few milliohms of current shunt.

My feedback resistors were so low that I added a buffer amp inside the loop just to drive that current, so's to not heat up the LT1028 transiently.

I can't complain about the price. Our noise level was 70x better than the competition.

Exactly. Surface mount has mostly killed the hybrid business, and I suspect a chunk of monolithic stuff. Apex used to be hybrids and high-voltage CMOS, and now they put parts on boards.

Or packaged parts on board!

LTZ1000 remains unique, if you need it. A few decades old and still unbeaten.

At one point they made the only 20bit SAR ADC (LTC2378-20).

Maybe LT5400?

LTC now makes a 24 bit SAR ADC. I'm dubious. It sounds like that are doing some heavy noise dithering or something.

Yes I was so unimpressed that I forgot about it. Trying to match TI with their "32-bit" ADC perhaps.

The 20bit one looked pretty good though.

Why would that make you dubious?

For direct-writing, shaped beam electron microfabricators are a bit faster. Bell Labs made one, and sold five to European Silicon Structures (ES2) at A ix-en-Provence. They were five times slower than Bell Labs had claimed (or so I was told last year) and the company eventually got restructured to dea l with the consequences of that. Cambridge Instruments used them to make a couple of special purpose chips - of which we were only going to buy 500 pe r year.

At the time Cambridge Instruments was working on their own shaped beam elec tron microfabricator - I was in charge of the electronic hardware - based o n a system that had been sold to us by Thompson-CFS as a pre-production pro totype, which was in fact a rather crude proof-of-principle machine.

It took us about a year to work out exactly how much work we'd have to do t o get it working, and when we did the project got cancelled - the company w ould have been doing nothing else for 18 months, which wasn't a practical o ption - so instead of spending 3.75 million UK pounds on finishing the proj ect, we spent 3.75 million UK pounds on buying ourselves out of all the con tracts we'd signed.

We were developing the S.360 - the first fully computer-controlled scanning electron microscope at the the time - and got a year or so jump on the mar ket when we released it. Sadly, somebody economised on special purpose test gear in the last stages of the development, and we couldn't make them as f ast as the market would have liked for that year or so, and by the time we' d developed the extra test gear, the competition had caught up.

Interesting times. The ES2 chips were to go into the S.420 and S.440 scanni ng electron microscopes, which were just as computer-controoled as the S.36

E.g. any CMOS ADC or DAC or DDS that costs more than 25 dollars. Also many Bicmos parts. Use Digi-key, I'm sure you can find some. Wafer fab prices tend to be confidential, though there are probably order-of-magnitude figures on the net somewhere. No I won't google that for you. Anyway you can believe what you want, I won't mind.

Yes, true. But IME packaging was less than half, (they were complex chips with few IO pins).

True, though for a few thousand they will do it. That might only be one wafer.

Sort-of, but I don't get 0.1% matching almost for free with discrete resistors.

If you have a very large ratio of current density (large ratio of emitter area times large ratio of currents), that helps. Headroom permitting, you can also stack two bandgap cells so that the noise only goes up 3dB but you get 6dB gain from putting them in series (and you can reuse the current so it is almost "free", and you avoid the noise of gain-setting resistors of an amplifier, and the noise of the transistors in the amplifier). We didn't have any characterised zeners on the processes that I used (a disadvantage of IC design).

It is nicer if you don't need massive (maybe leaky, maybe microphonic) capacitors in the filtering. It is also nice if it settles within 10us of being disturbed or turned on. (e.g. in GSM cellphones it allows the quiet regulators to be disabled between bursts to save power. When waiting for a call, the radio turns on for maybe some milliseconds every

Pretty nearly nowadays--0.01% metal film resistors are a few cents apiece, probably less than it costs to place them on the board. Check out Digikey for the Panasonic ERA series, for instance. Plus you can get tiny thin-film resistor arrays for a penny or two, whose tempcos track quite a bit better than their absolute resistance. They come in a much wider range of values than on-chip resistors, too.

By the time you get enough bandgaps in series to make the noise competitive, you might as well use an LM329, or even two in series. Normal bandgaps are the pits for SNR, and by adding sections, you gain it back only linearly in power (i.e. by the square root in voltage terms).

But there are lots of super good capacitors available for under a penny. Good luck making a 1-Hz RC rolloff on-chip without a horrible noise penalty. I's easy in a built-up circuit. If you're using some nasty high-K ceramic and are worried about microphonics, you just rout out a little paddle on the board for the cap to sit on, so it's isolated from the board flex. Easy-peasy.

Different things are easy and hard in the two domains, as I said. In a built-up circuit you'd probably do that with a baby-scale loop, i.e. one that uses a fast TC till it settles to within some fraction of a percent, and then switches to a longer one to reject noise. I think Win and Paul talk about that someplace.

Cheers

Phil Hobbs