Wide temperature range voltage reference design

[d123]

Hi,

I don't know where to post this, it's not DIY, but I'm not an IC designer, 'though I am happy with this design idea, an idea that I think could be enormously improved and simplified by some-one who knows what they're doing.

I've been trying to design a simple flexible voltage reference that is stable over -50 to 150ºC, and although this is fairly pathetic (too much current, too many parts, probably other issues I can't see), I am happy with the result, so here it is:

A centigrade temperature sensor in a Burr-Brown/TI app note called DIODE-BASED TEMPERATURE MEASUREMENT (sboa019) is the basis for the circuit, I added a buffer to carry the transistor Vbe drops over temp., both fed into a non-inverting summing amplifier, followed by another buffer for the output.

It's a pedestrian idea, sum two opposing dV/dTemps, then divide them, and the product will be a constant value.

It theoretically maintains 1.23V to 1.24V over -50 to 150ºC with 5ppm resistors, and 1.23V from 0 to 100ºC, it's only a simulation, but I've had worse ideas. Just want to share the concept, I don't think it's great or groundbreaking, I know it's nothing special, just an idea.

 

[dick_freebird]

First, there are (and I have designed & seen installed to
production) precision voltage references that are
rated over at least -55C to 125C. What I do not see, is a
quantitative spec on "stable". This of course varies by the
application; a power supply reference is good at 1% but
if you imagine you're going to make a 16-bit data acquisition
lineup, well, welcome to the Unobtanium Shopping Network.

Not that my parts are that much of a bargain. They cost
more than any of my first four cars. You could call them
unobtanium-plated.

The devil is in some details. For example, all of those
perfectly matched and temperature-insensitive current
sources biasing the reference elements. Hand-whittled
from billet unobtanium by unionized asthmatic gnomes
of a protected class with fully vested guaranteed-benefit
pensions, may as well be.

If it were my system I think I'd think a lot about figuring
the system transfer function of {output quantity} to
VREF, cal-map true VREF vs temp with good testing
(presuming that a population acts the same, even if not
ideally or to industrial-temp-range spec window) and
then de-embed, by inverse of transfer function, that
cal-curve using some remote big brain.

Or if it's a power supply, just blow out the high temp
spec a wee bit and see if it costs you anything by the
customer's lights.

If the temp sensor IC can be counted upon (the bandgap
reference's curvature likewise) or you built in enough
post-assembly tweakability to cover the components'
parametric dispersion, it's not a bad idea. Of course many
demanding-environment applications don't like trimpots
much, either....

There are fancy-schmancy references out there that
do the cal-map thing internally, but they are more after
super precision across limited temp range, than flatness
across extended temp range. And they may be pretty
"needy" on the production test floor because they have
to get that cal-map stuffed into them based on real per-
unit testing (and that's even-better gear, run across
temp in fine steps, yadda yadda). reference bandgap,
temp sensor, an ADC on the temp sensor, a lookup
table of nonvolatile (or OTP) memory and a DAC to bend
the reference right. I wouldn't call it elegant but it's the
only way I've seen to hyper-precision voltage reference.

 

[d123]

Hi,

...So it's not a threat to the quality of floating gate references then!

Some of what you say half goes over my head, but sort of understand most of it. Need to find out what "cal-map" is, for example. I know one tweakability method is zener-zapping, and fuses, resistors, etc., the circuit is imperfect enough without throwing trimpots in too.

Last thing you describe for precision reference sounds like a machine, in a good sense. Anyway, for some reason the "helped me" button isn't appearing, so, thanks.

 

[dick_freebird]

Floating-gate references can be nudged very fine-grained
to an ideal measured output. However floating-gate devices
have (even in digital, agressively written and caring only about
making a 1/0 decision) a read lifetime strongly degraded by
temperature. I have never seen one advertised (doesn't mean
they don't exist, but my interest-space has other deleterious
actors as well) outside commercial temp range.

Zener zap, fuse link, laser trim, all of these are applied to
piece-part production. The more aspects you want to trim,
the more cost you embed and the more you depend on the
stability of individual devices (once trimmed, it's all up to
the drifts in transistor Vbe, VT match and reliability expectations
in this respect are highly variable - in many cases there's a
lot of chest-bumping behind the scenes between what the
marketing folks want to sell, what the designers say they need
to do it, what the technology development folks are willing to
promise and who the reliability guys think is trying to put a fast
one over.

By "cal-map" I mean a calibration "map" which has some index
(e.g. digitized temperature, not limited to one) and some
output (likely converter back to have some analog effect)
and is written based on multiple test points' result. A "map"
stored as writable digital bits rather than ablated thin film
resistor material or fried zeners or opened poly resistor necks.
This is perhaps a fine and arbitrary distinction, I guess. More
about implementation and complexity. You wouldn't do an
ADC, 16x16 memory, DAC in a 10um bipolar process but for
modern CMOS the area probably fits under a bond pad, the
transistors are poorly matched so range*resolution requires
more bits than a purpose-designed analog flow (ADI used to
develop a flow for a product; good luck at TSMC) and in the
end you have to use a scheme that the foundry you have,
supports) and meanwhile expectations creep higher.

A bandgap (or zener) reference likely will have two trims
(tempco, and then voltage to take out any movement that
the tempco trim induced). You could consider adding VIN
as a means to enhance DC PSRR. Or whatever. The more
you intend to touch and the finer you intend to slice it,
the more (perhaps geometrically) your per-unit test costs
blow out (not to mention, more fussiness of part spec
leads directly to more expensive equipment and more
test engineering to make its results consistent, credible
and supporting final test outcomes).

 

[dick_freebird]

I just ran across what you could consider a challenge,
and a solution.

https://www.linear.com/docs/54315
Linear Tech's new LT6658 reference looks like a do-all
part, available in -40C - 125C spec grades (but they
do play the usual game with a looser spec as the price
of a wider rating-range). If you can stand 0.215%
tolerance out to 125C (the LT6658AH - probably much
more $ than the advertised commercial-range LT6658 at
$2.35@1Kpcs) then this is worth a look.

I and probably anyone else who's done a vref product
design would consider it quite the challenge to match
(especially, done on what looks like "40V" bipolar
technology in the classical way). I'd almost buy one
just to peer at the layout styles under my microscope.
I'm sure there's some real art inside.

 

[d123]

Hi,

Thanks, I'll have a read of that, I like trying to understand the block diagrams, and slowly learning to interpret the detailed simplified schematics when included, great learning tools. That's an impressive %, 0.215... Whilst doing my unprofessional studying I came across the LT1407 (by memory, may have part number slightly wrong), don't remember it being as accurate as that.

Good to hear a professional take on floating gate, looks like the Holy Grail of references as they have specs of less than 2mV inaccuracy, until it's explained a bit better.

Design art, and chip art maybe?!

Thanks for all the interesting and informative advice, it's great to hear the point of view and knowledge of an experienced designer, much appreciated. "Helped me lots."

- - - Updated - - -

Correction, the 1407, if it's that part number, is a temperature sensor...

The LT6658 is a very simple design, in a positive sense (not like my circuit which needs too many parts), one bandgap, one op amp, one feedback pin, and 0.06% change in reference over that wide temp range. Thanks.