Re: Pete Lefferts LED current source

On Tue, 28 Oct 2008 15:29:18 -0500, Phil Hobbs
<pcdhSpamMeSenseless@xxxxxxxxxxxxxxxxxx> wrote:

ggherold@xxxxxxxxx wrote:
On Oct 26, 3:54 pm, "Howard Swain" <hsw...@xxxxxxxxxxxxx> wrote:
<ggher...@xxxxxxxxx> wrote in messagenews:31c70ce5-55cb-4cda-9ac2-161747ed6911@xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
Phil Hobbs wrote:
ggher...@xxxxxxxxx wrote:
On Oct 23, 5:06 pm, Phil Hobbs
<pcdhSpamMeSensel...@xxxxxxxxxxxxxxxxxx> wrote:
ggher...@xxxxxxxxx wrote:
On Oct 21, 8:07 pm, Phil Hobbs
<pcdhSpamMeSensel...@xxxxxxxxxxxxxxxxxx> wrote:
I've used theLED/BJT follower trick often--I first saw it in probably
1980, in an app note describing a really quiet battery-powered mic
preamp. Noise is usually much more important to me than small amounts
of drift, so I like using forward-biased LEDs for voltage references.
You get a free pilot light out of the deal, too.
Forward-biased diodes have a noise temperature of 150K at room
temperature, so if you drive it reasonably hard, you can make a really
quiet voltage reference this way. (I'm sure there'll be some raised
eyebrows in this group over a 'voltage reference' with 150 uV/K drift.
Take it up with Maxwell's Demon.)
Cheers,
Phil Hobbs- Hide quoted text -
Phil Hobbs wrote,
" Noise is usually much more important to me than small amounts
of drift, so I like using forward-biased LEDs for voltage references.
You get a free pilot light out of the deal, too.
Forward-biased diodes have a noise temperature of 150K at room
temperature, so if you drive it reasonably hard, you can make a really
quiet voltage reference this way. (I'm sure there'll be some raised
eyebrows in this group over a 'voltage reference' with 150 uV/K
drift.
Take it up with Maxwell's Demon.)"
Darn, You mean I can make a quiter voltage reference? To get a 1nV/
rtHz voltage reference I've been filtering the piss out of the
standard IC voltage references (buried Zeners) These have a great
TempCo's but 100nV/rtHz or greater noise.
Say, I tried to "derive" the stated 150uV/K drift but I came up with
numbers an order of magnitude bigger. Is this the difference between
the drift of the BJT andLED? The Maxwell's Demon comment makes me
think it is something more fundemental.
George Herold
PS. thanks all for the nice discussion. I think I "got" about half of
it. The rest can be mulled over at my leisure
G
Yes, you can make a quieter low-voltage reference--a good 20 dB quieter
than a bandgap, if you're willing to spend some current. The noise
model of a PN junction is full shot noise (sqrt(2eI) in 1 Hz) in
parallel with the differential resistance of the junction, which for an
ideal diode (e.g. a diode-connected transistor) is kT/(eI). When you
combine those two formulas, and compare it with the Johnson noise
formula, you get the useful result that T_noise = T_j/2.
Real diodes are a bit noisier than this because although they're
reasonably exponential, the constant in the exponent isn't kT/e (25 mV
at room temperature) but a bit higher, 35-50 mV in most small signal
devices. That means roughly that if you spend a milliamp in the
transistor and a milliamp in theLED, you can make a 1V reference whose
noise is on the order of 1.5 nV in 1 Hz, including both the diode and
transistor noise.
Bandgaps work by adding a proportional-to-absolute-temperature (PTAT)
voltage to a V_BE drop, and adjusting the total voltage until the drift
cancels, which happens at about 1.22 V, the zero-temperature band gap of
silicon (hence the name). (IC guys like Jim and Walt know lots more
about bandgaps, including stuff like higher-order corrections for the
curvature of the V(T) curve, but I'm a noise guy.) The PTAT voltage
comes from the DeltaV_BE between two transistors running at roughly 10X
different collector current densities, which comes out to roughly 60 mV
at 300K. That DeltaV_BE has to be amplified by about 10 times before
being added to the V_BE drop to make the 1.22V output voltage, and
that's where the problem is.
The absolute noise level on the 60 mV is reasonably low, assuming that
the collector currents are the same (i.e. the current density ratio
comes from scaling the device area rather than the collector current).
The V_BE is quiet too, but unfortunately you have to apply about 20 dB
gain to the DeltaV_BE portion, which makes it very noisy. Bandgap
designers also work under a lot of pressure to reduce operating current,
which doesn't help the noise one bit.
On the other hand, you can get far better temperature stability with a
bandgap, and if you really need to, you can combine 10 of them to reduce
the noise.
The 150 uV/K number is a SWAG for how closely you can expect the two
temperature coefficients to track each other. You can adjust it a bit
by changing the diode current--lower current equals higher drift. In
general theLED/emitter follower trick is great for most things except
A/D references.
I haven't done a 1/f noise comparison between the two kinds of devices,
but V_BEs generally have very very low 1/f corners.
Cheers,
Phil Hobbs- Hide quoted text -
- Show quoted text -
Thanks Phil! I now have a new circuit to try some day. I wish there
was a means to "buy you a beer" via the web.
The PTAT and diode discussion "speaks" to me. I'm using a diode
connected transistor (2N3094) as a temperature sensor. This has two
modes of operation. First one can run a constant current (10uA typ.)
through the diode and measure the voltage drop. You calibrate it at
a few known temperatures and interpolate in between. You can then
also look at the forward voltage drop at two different currents (at
the same temperature), the "slope" is given by the thermal voltage kT/
e. The sensor is almost self calibrating. For the 2N3904 I find the
error is a little less than 1%. With the predicted temperature one or
two degrees higher than what is measured by some other means. I
assume this is the non-ideality factor of the transitor. I wonder
what causes this non-ideality and if there is a different transistor
(or diode) that is more ideal?
George Herold
A 1% error might be coming from beta nonlinearity--it's I_C that's
exponential, not I_E. You might get better results with something like
an MPSA18, which has a huge beta, or a 2N5086, which IIRC has amazing
beta linearity. It might also be self-heating or the extrinsic emitter
resistance, if you're running too much current. The sweet spot for that
sort of measurement is around 100 uA or a little bit below. There are
circuit things you can do as well, e.g. drive the base from a follower
and measure V_BE rather than V_CE.
Cheers,
Phil Hobbs
Oh I'm only using the transistor for it's base emitter diode. There
is no gain in the circuit. I measure the forward voltage drop with
currents from 10nA to 1mA. When plotted semi-log you get a nice
straight line of slope kT/e. You can do the same for the base
collector voltage but the slope is further from ideal. And as you
said previously, small signal diodes show even larger slopes. I've
heard the base emitter junction is more aburpt and thought this might
have something to do with the "better" behavior of the b-e junction.
Perhaps less carrier recombination in the depletion region, but now
I'm speaking of things of which I know very little. I would like to
know what transistor specification might lead to more ideal V_BE
behavior and I'll try some of the high beta's. Though I also wonder
if higher speed transistors might be better. Oh, the emitter
resistance is something I thought about only recently, at 10nA the
"diode" has 25 Meg of source impedance, I must have had an opamp
buffer in the circuit beofre the voltmeter else I would have noticed
non-exponential behavior at the low currents. I hoping to use these
down to 77K, but the plastic packages may not survive repeated trips
into the liquid nitrogen and I'll have to "fall back" to just using
the glass encapsulated diodes which seem to survive just fine. So
finding a better transistor is not an issue until I know that the
packages will work.
Thanks for the suggestion,
George Herold
Say you have I = Is exp (qV/nkT), where n is the ideality constant.

See the GE Transistor Manual, seventh edition (1964), pp. 439-442.
They mention that (for your kind of currents) n is about 2 for gold doped
and can approach 1 for non-gold doped. The "impurity gradient" is also
mentioned, but I don't see that they explain its effect.

Many years ago I did measurements similar to yours and found that
diodes I'd expect to be gold doped (such as 1N4000 rectifiers and
1N914-type fast switching) did have an n close to 2. The only Si
pn diodes I found with an n close to 1 were varactors. Of course,
hot carrier diodes had n close1, also. As I recall, old-fashioned
"fast switching" transistors (perhaps 2N2369) also seemed to be
gold doped. I'd expect modern microwave transistors with ft > 5GHz
not to be gold doped, but I've never measured them.

Does your "error is a little less than 1%" mean that you are measuring n of 1.01?
If so, I'd be surprised if you could do closer to 1 than that.

--
Regards,
Howard- Hide quoted text -

- Show quoted text -

"Does your "error is a little less than 1%" mean that you are
measuring n of 1.01?
If so, I'd be surprised if you could do closer to 1 than that."

Yes, Exactly. n < 1.01 for V_BE of 2N3904 transistor. Hmm, 1/(n-1)
is suspiciously close to the Beta of the transistor. I definitely
have to order some of the high Beta tranistors recommended by Phil H.
I'll admit I don't really understand transistors and what factors
determine the current gain. In fact it's only recently that I feel
like I'm beginning to understand the humble diode.

George Herold



Laser noise cancellers can get up to 70 dB of cancellation (1 part in 3000
current error) with unmatched transistors running at very different current
densities, which shows that BJTs follow the Ebers-Moll law very closely,
until the extrinsic resistances become important. If your error is as much
as 1%, you're not getting the performance available from your 2N3904s.

Ebers-Moll is not a law but merely model of how a transistor works.
Spice programs have moved to Gummel-Poon models for BJTs and have gone
through some 4 levels of MOS models.


If you connect them as diodes (base and collector shorted), they actually run
as normally-biased transistors--in fact that's how the noise canceller uses
them. If your error were 0.1%, I'd begin to believe that it might be the
transistors, but 1% is _way_ too large. That's definitely a circuit problem
or perhaps a thermal leak--remember that the transistor leads are roughly
4000 times more thermally conductive than the plastic case.

National has a useful temperature control handbook that talks a lot about the
problem of temperature measurement, and admits in so many words that IC
temperature sensors (and packaged transistors) measure the temperature of
their leads.

BJTs are really amazing devices--they follow their simple models essentially
exactly, unlike FETs.

Cheers,

Phil Hobbs

Besides, ggherold should also try diode connected transistors, it
would extend the plots by an order of magnitude or two, with better
conformance. It is too bad that the V(be) reverse bias breakdown does
not change, or this would be a more popular configuration.

.