Switching transistors, current sources, nonidealties and noise
Good evening,
I have recently been looking into BJT's and their switching properties.
Because a time-to-amplitude converter does similar things as I would like to,
I have been looking in what people do with those. First thing that strikes
me as kind of odd is that almost all designs I have seen use some general
purpose transistor (like 2N2222, 2N3904, BC848...). The only two exceptions
seems to be Guide Technology, who use an UPA806T (RF npn pair) for the
diff-pair current switch[1] in their TAC and a group at Oulu University[2].
But even Guide Technology uses an 2N3904 for the current source.
Having put the circuit through Spice, I see that the current through
the tail fluctates violently during the time when the current switches
from one transistor of the pair to the other. The reason for this seems
to be that the f_t of the current source transistor is too low to compensate.
Trying to replace the current source with an RF transitor like BFU520
that has an f_t of 10GHz helps to dampen these fluctuations by a factor of 2,
but they are still there.
Why do people use general purpose transistors in these places, even
though RF transistors definitly improve switching behaviour?
I dimply remember that someone said/wrote once, that RF transistors have
a higher noise. But if I look at the datasheet, the quoted noise figure
for the BFU520 is <1.6dB while the noise figure of the 2N3904 is 2dB best case.
As I still do not really know how to read single transistor datasheets,
I am pretty sure I missed something fundamental there.
Attila Kinali
[1] "Time Interval Analyzer Having Interpolator With Constant Current Capacitor Control", US Patent 6'091'671
[2] "Wide-Range Time-To-Digital Converter With 1ps Single-Shot Precision",
by Keränen, Määttä, Kostamovaara, 2011
--
Malek's Law:
Any simple idea will be worded in the most complicated way.
Hi
Leakage
Bob
On Jun 19, 2016, at 1:23 PM, Attila Kinali attila@kinali.ch wrote:
Good evening,
I have recently been looking into BJT's and their switching properties.
Because a time-to-amplitude converter does similar things as I would like to,
I have been looking in what people do with those. First thing that strikes
me as kind of odd is that almost all designs I have seen use some general
purpose transistor (like 2N2222, 2N3904, BC848...). The only two exceptions
seems to be Guide Technology, who use an UPA806T (RF npn pair) for the
diff-pair current switch[1] in their TAC and a group at Oulu University[2].
But even Guide Technology uses an 2N3904 for the current source.
Having put the circuit through Spice, I see that the current through
the tail fluctates violently during the time when the current switches
from one transistor of the pair to the other. The reason for this seems
to be that the f_t of the current source transistor is too low to compensate.
Trying to replace the current source with an RF transitor like BFU520
that has an f_t of 10GHz helps to dampen these fluctuations by a factor of 2,
but they are still there.
Why do people use general purpose transistors in these places, even
though RF transistors definitly improve switching behaviour?
I dimply remember that someone said/wrote once, that RF transistors have
a higher noise. But if I look at the datasheet, the quoted noise figure
for the BFU520 is <1.6dB while the noise figure of the 2N3904 is 2dB best case.
As I still do not really know how to read single transistor datasheets,
I am pretty sure I missed something fundamental there.
Attila Kinali
[1] "Time Interval Analyzer Having Interpolator With Constant Current Capacitor Control", US Patent 6'091'671
[2] "Wide-Range Time-To-Digital Converter With 1ps Single-Shot Precision",
by Keränen, Määttä, Kostamovaara, 2011
--
Malek's Law:
Any simple idea will be worded in the most complicated way.
time-nuts mailing list -- time-nuts@febo.com
To unsubscribe, go to https://www.febo.com/cgi-bin/mailman/listinfo/time-nuts
and follow the instructions there.
On 6/19/16 10:23 AM, Attila Kinali wrote:
Good evening,
I have recently been looking into BJT's and their switching properties.
Because a time-to-amplitude converter does similar things as I would like to,
I have been looking in what people do with those. First thing that strikes
me as kind of odd is that almost all designs I have seen use some general
purpose transistor (like 2N2222, 2N3904, BC848...). The only two exceptions
seems to be Guide Technology, who use an UPA806T (RF npn pair) for the
diff-pair current switch[1] in their TAC and a group at Oulu University[2].
But even Guide Technology uses an 2N3904 for the current source.
Having put the circuit through Spice, I see that the current through
the tail fluctates violently during the time when the current switches
from one transistor of the pair to the other. The reason for this seems
to be that the f_t of the current source transistor is too low to compensate.
Trying to replace the current source with an RF transitor like BFU520
that has an f_t of 10GHz helps to dampen these fluctuations by a factor of 2,
but they are still there.
Why do people use general purpose transistors in these places, even
though RF transistors definitly improve switching behaviour?
I dimply remember that someone said/wrote once, that RF transistors have
a higher noise. But if I look at the datasheet, the quoted noise figure
for the BFU520 is <1.6dB while the noise figure of the 2N3904 is 2dB best case.
As I still do not really know how to read single transistor datasheets,
I am pretty sure I missed something fundamental there.
I'll bet the noise contribution is trickier to figure out. There's
current noise and voltage noise, for one thing. A few months back, we
were designing (and building) a low noise amplifier for 5-30 MHz.
There's a whole trade between microwave parts (specified in terms of NF
with 50 ohm source, typically down to the "DC" frequency of 50MHz) and
Op-Amps (which have noise voltage and current plots in the data sheet)
and discrete devices of one sort or another (which have scant noise
information).
There's also huge differences among specific mfrs and lots for the same
JEDEC number (e.g. 2n2222 or 2n3904) and noise, since to use the JEDEC
number, you just have to meet the requirement.. after that, you could be
orders of magnitude better or just squeak by. There's a very low
leakage JFET (that I can't remember the number) where there's one
company that makes "really, really good" ones, and the rest are "meet
the data sheet". Nuclear instrumentation amplifier builders want only
the "special" ones.
There's plenty of low noise amplifier designs out there and the
descriptions typically assert "we selected this device because it had
good noise properties", but then, typically, do not explain why that one
worked, and the other half dozen that are superficially similar (on the
data sheet or in terms of internal construction) didn't. I suspect
that in some cases, it's what they had in stock, they built it, it
worked, and then there was no more to be done.
There's also the "other requirements" aspect: maybe you need really low
noise, but you also need strong signal handling at the same time, so
you're running your device with a ton of drain/collector current so it
doesn't saturate. So you had to pick a device that would take that.
Or, you needed a device that could tolerate 30V, because you're running
off a 24V supply.
The SPICE model probably is the "nominal device"...
Ultimately, you have to get some sample devices and measure them (no
trivial matter in itself).
Why do people use general purpose transistors in these places, even
though RF transistors definitly improve switching behaviour?
Commercial designs do use RF transistors but only old ones are
documented.
The Tektronix 7A11 uses 2 GHz PNPs and 1 GHz NPNs but its design is
unusual since it can integrate positive or negative time and while the
transistors use emitter switching, they are not configured as
differential pairs.
The Tektronix 2440 uses a 1.2 GHz NPN differential pair for the fast
ramp switching and a 2N3906 differential pair for the slow ramp
switching.
Having put the circuit through Spice, I see that the current through
the tail fluctates violently during the time when the current switches
from one transistor of the pair to the other. The reason for this seems
to be that the f_t of the current source transistor is too low to compensate.
Trying to replace the current source with an RF transitor like BFU520
that has an f_t of 10GHz helps to dampen these fluctuations by a factor of 2,
but they are still there.
Of interest in these designs is that they do not use separate
transistor current sources where fast switching is involved; the
differential pairs do double duty and the tail current is set by a
resistor to a separately decoupled bias supply.
On Sunday, June 19, 2016 07:23:09 PM Attila Kinali wrote:
Good evening,
I have recently been looking into BJT's and their switching properties.
Because a time-to-amplitude converter does similar things as I would like
to, I have been looking in what people do with those. First thing that
strikes me as kind of odd is that almost all designs I have seen use some
general purpose transistor (like 2N2222, 2N3904, BC848...). The only two
exceptions seems to be Guide Technology, who use an UPA806T (RF npn pair)
for the diff-pair current switch[1] in their TAC and a group at Oulu
University[2]. But even Guide Technology uses an 2N3904 for the current
source.
Having put the circuit through Spice, I see that the current through
the tail fluctates violently during the time when the current switches
from one transistor of the pair to the other. The reason for this seems
to be that the f_t of the current source transistor is too low to
compensate. Trying to replace the current source with an RF transitor like
BFU520 that has an f_t of 10GHz helps to dampen these fluctuations by a
factor of 2, but they are still there.
Why do people use general purpose transistors in these places, even
though RF transistors definitly improve switching behaviour?
I dimply remember that someone said/wrote once, that RF transistors have
a higher noise. But if I look at the datasheet, the quoted noise figure
for the BFU520 is <1.6dB while the noise figure of the 2N3904 is 2dB best
case. As I still do not really know how to read single transistor
datasheets, I am pretty sure I missed something fundamental there.
Attila Kinali
[1] "Time Interval Analyzer Having Interpolator With Constant Current
Capacitor Control", US Patent 6'091'671
[2] "Wide-Range Time-To-Digital Converter With 1ps Single-Shot Precision",
by Keränen, Määttä, Kostamovaara, 2011
Smaller Early effect at operating voltage for 2N3904 than lower voltage high
frequency transistor and hence lower current modulation?
When measuring the delay of a 2 stage synchroniser,the delay range of interest
is 1-2 synchroniser clock periods and nonlinearities within the first clock
period are of little interest as long as the effect is repeatable and has
settled out within 1 clock period of the transition, the effect is merely an
offset and thus of little consequence.
Bruce
Attila wrote:
Having put the circuit through Spice, I see that the current through
the tail fluctates violently during the time when the current switches
from one transistor of the pair to the other. The reason for this seems
to be that the f_t of the current source transistor is too low to compensate.
Trying to replace the current source with an RF transitor like BFU520
that has an f_t of 10GHz helps to dampen these fluctuations by a factor of 2,
but they are still there.
The transation frequency of the current source transistor is part of the
cause, but the primary cause is generally the capacitance of the CS
output node to ground. Some designers put an inductor in series with
the output, but I have never found this to be very effective [except in
poorly-designed simulations] due to the self-capacitance of the
inductor. Much better, IME, is to add a cascode device to the current
source. (See attached images.) This has the added benefit of
increasing the output resistance. This increase can be very substantial
(several orders of magnitude) if you use a FET cascode device as shown.
Why do people use general purpose transistors in these places, even
though RF transistors definitly improve switching behaviour?
I dimply remember that someone said/wrote once, that RF transistors have
a higher noise. But if I look at the datasheet, the quoted noise figure
for the BFU520 is <1.6dB while the noise figure of the 2N3904 is 2dB best case.
I, for one, have said this, but you are not remembering the whole point.
RF transistors are generally considerably noisier AT BASEBAND than GP
transistors, both because their geometries are inherently noisier and
because they have much higher flicker noise corner frequencies
(usually 10kHz to some MHz for RF transistors, compared to 10Hz-1kHz for
GP transistors). One might think that this would not matter at RF, but
the flicker noise modulates the bias of the transistor (and sometimes
other circuit elements), leading to both simple noise modulation as well
as phase modulation. RF transistors are not specified for noise at
baseband.
Referring to the attached images:
Circuit 1 (files ...01a and ...02a) is an LTspice simulation of a
cascoded ~3mA current source running into a node that shifts up and down
by 1v at 10MHz with rise and fall times of 1nS (vaguely simulating the
emitter node of a BJT differential pair switching at 10MHz). The green
trace is the current source output (drain of J1), the red trace is the
collector of Q1, and the cyan trace is the voltage forced at the CS
output through its internal 50 ohm resistance. Only the positive
transition is shown -- the negative transition is substantially the
same. The current increases by ~25% during the 1nS transitions due to
the output capacitance of the FET. NOTE: There will also be stray
capacitance at the output node, which will make this worse in practice.
Circuit 3 (files ...03a and ...03b) adds an actual differential pair
with unbalanced 2v p-p drive. As expected, the simulation shows
somewhat larger and somewhat longer perturbations at the transitions.
General comments:
Like any simulation, one needs to understand the limitations of both the
models used and the circuit choices made by the designer. I cannot
possibly write enough here to cover the particulars of this simulation
adequately. While my experience tells me that my choices are reasonable
and that the simulation is generally valid, there are still details to
be mindful of. You are warned!
I make no claim that these are the best devices to use for a 10MHz TAC,
although IME they are reasonable choices.
I modeled the ...01a circuit using a BFR90A BJT as the cascode device,
and the simulation showed that the current spikes were reduced by about
50%. However, my experience tells me that this would not hold in
practice. The simulation also showed that the DC output resistance fell
from nearly 300Mohm with the FET to only 1.4Mohm -- which experience
tells me is in the ballpark of how the real circuits would perform.
Best regards,
Charles
Moin,
Thanks everyone for the answers!
On Mon, 20 Jun 2016 01:45:24 -0400
Charles Steinmetz csteinmetz@yandex.com wrote:
The transation frequency of the current source transistor is part of the
cause, but the primary cause is generally the capacitance of the CS
output node to ground. Some designers put an inductor in series with
the output, but I have never found this to be very effective [except in
poorly-designed simulations] due to the self-capacitance of the
inductor. Much better, IME, is to add a cascode device to the current
source. (See attached images.) This has the added benefit of
increasing the output resistance. This increase can be very substantial
(several orders of magnitude) if you use a FET cascode device as shown.
I simulated a couple of circuits, with very different results.
First thing that struck me was, that it is neigh impossible to
make cascode circuits stable when using RF transistors. And even
if I managed to do that, small changes in resistor values would
imediatly make it oscillate again, or degrade performance severely.
Same goes for using Darlington circuits (which I tried in order to
minimize the effects of beta variation).
The best results I got was with the attached circuit. Ie using
a classical opamp based npn current source, but using an emitter
follower between transistor and opamp in order to enhance high
frequency (aka transient) performance. R29 is there to load Q7
and to prevent it from going into saturation. R30 is needed for
stabilizing the circuit (I do not exactly understand what the
mechanics of the oscillations are, when R30 is removed, if someone
knows, please tell me). The voltage divider R30/R31 helps to keep
the opamp output away from the lower power rail. If stability is still
an issue, a 5-10pF capacitor should be added from the output of the
opamp to the inverting input (degrades frequency response below 1MHz slightly).
The simulation output shows the current through the (zero) voltage source
at the tail of the differential pair. The I(V11) curve is the circuit as
shown and the I(V7) curve is the same circuit with the two BFU520 replaced
by 2N3904. As can be seen, the transient of the 2N3904 is several times
larger than the one of the BFU520 and lasts for about three times as long.
I have not done any analysis of the temperature stability, yet.
My guess would be that is dominated by the input offset voltage
temperature coefficient of the opamp. But I have no calculations
to prove it.
Noise analysis would be interesting, but I doubt there is enough
data available to actually get some meaningfull results out of it.
Why do people use general purpose transistors in these places, even
though RF transistors definitly improve switching behaviour?
I dimply remember that someone said/wrote once, that RF transistors have
a higher noise. But if I look at the datasheet, the quoted noise figure
for the BFU520 is <1.6dB while the noise figure of the 2N3904 is 2dB best case.
I, for one, have said this, but you are not remembering the whole point.
RF transistors are generally considerably noisier AT BASEBAND than GP
transistors, both because their geometries are inherently noisier and
because they have much higher flicker noise corner frequencies
(usually 10kHz to some MHz for RF transistors, compared to 10Hz-1kHz for
GP transistors). One might think that this would not matter at RF, but
the flicker noise modulates the bias of the transistor (and sometimes
other circuit elements), leading to both simple noise modulation as well
as phase modulation. RF transistors are not specified for noise at
baseband.
Hmm.. if the flicker noise corner frequency would be in the few 10kHz to
100kHz range, then I would not be worried. The opamp's control loop
should "kill" anything below ~100kHz and dampen quite a bit up to 1-2MHz.
I would even suspect that in the <10kHz range, the noise of the opamp would
dominate the noise of the transistors.
I modeled the ...01a circuit using a BFR90A BJT as the cascode device,
and the simulation showed that the current spikes were reduced by about
50%. However, my experience tells me that this would not hold in
practice.
Do you know where the discrepancy between simulation and reality comes from?
Attila Kinali
--
It is upon moral qualities that a society is ultimately founded. All
the prosperity and technological sophistication in the world is of no
use without that foundation.
-- Miss Matheson, The Diamond Age, Neil Stephenson
Hi
On Jul 1, 2016, at 7:56 AM, Attila Kinali attila@kinali.ch wrote:
Moin,
Thanks everyone for the answers!
On Mon, 20 Jun 2016 01:45:24 -0400
Charles Steinmetz csteinmetz@yandex.com wrote:
The transation frequency of the current source transistor is part of the
cause, but the primary cause is generally the capacitance of the CS
output node to ground. Some designers put an inductor in series with
the output, but I have never found this to be very effective [except in
poorly-designed simulations] due to the self-capacitance of the
inductor. Much better, IME, is to add a cascode device to the current
source. (See attached images.) This has the added benefit of
increasing the output resistance. This increase can be very substantial
(several orders of magnitude) if you use a FET cascode device as shown.
I simulated a couple of circuits, with very different results.
First thing that struck me was, that it is neigh impossible to
make cascode circuits stable when using RF transistors.
Real cascode circuits can be built with RF transistors. They also can be simulated.
Simulating them with the “standard” models is a PIA. The issue is that the inductance
of the package is not de-embedded from the test “socket” as carefully as it might be.
There is also the somewhat non-intuitive need to stick a low value resistor in the base.
Done properly, they are very reproducible and reasonably insensitive to load.
Bob
And even
if I managed to do that, small changes in resistor values would
imediatly make it oscillate again, or degrade performance severely.
Same goes for using Darlington circuits (which I tried in order to
minimize the effects of beta variation).
The best results I got was with the attached circuit. Ie using
a classical opamp based npn current source, but using an emitter
follower between transistor and opamp in order to enhance high
frequency (aka transient) performance. R29 is there to load Q7
and to prevent it from going into saturation. R30 is needed for
stabilizing the circuit (I do not exactly understand what the
mechanics of the oscillations are, when R30 is removed, if someone
knows, please tell me). The voltage divider R30/R31 helps to keep
the opamp output away from the lower power rail. If stability is still
an issue, a 5-10pF capacitor should be added from the output of the
opamp to the inverting input (degrades frequency response below 1MHz slightly).
The simulation output shows the current through the (zero) voltage source
at the tail of the differential pair. The I(V11) curve is the circuit as
shown and the I(V7) curve is the same circuit with the two BFU520 replaced
by 2N3904. As can be seen, the transient of the 2N3904 is several times
larger than the one of the BFU520 and lasts for about three times as long.
I have not done any analysis of the temperature stability, yet.
My guess would be that is dominated by the input offset voltage
temperature coefficient of the opamp. But I have no calculations
to prove it.
Noise analysis would be interesting, but I doubt there is enough
data available to actually get some meaningfull results out of it.
Why do people use general purpose transistors in these places, even
though RF transistors definitly improve switching behaviour?
I dimply remember that someone said/wrote once, that RF transistors have
a higher noise. But if I look at the datasheet, the quoted noise figure
for the BFU520 is <1.6dB while the noise figure of the 2N3904 is 2dB best case.
I, for one, have said this, but you are not remembering the whole point.
RF transistors are generally considerably noisier AT BASEBAND than GP
transistors, both because their geometries are inherently noisier and
because they have much higher flicker noise corner frequencies
(usually 10kHz to some MHz for RF transistors, compared to 10Hz-1kHz for
GP transistors). One might think that this would not matter at RF, but
the flicker noise modulates the bias of the transistor (and sometimes
other circuit elements), leading to both simple noise modulation as well
as phase modulation. RF transistors are not specified for noise at
baseband.
Hmm.. if the flicker noise corner frequency would be in the few 10kHz to
100kHz range, then I would not be worried. The opamp's control loop
should "kill" anything below ~100kHz and dampen quite a bit up to 1-2MHz.
I would even suspect that in the <10kHz range, the noise of the opamp would
dominate the noise of the transistors.
I modeled the ...01a circuit using a BFR90A BJT as the cascode device,
and the simulation showed that the current spikes were reduced by about
50%. However, my experience tells me that this would not hold in
practice.
Do you know where the discrepancy between simulation and reality comes from?
Attila Kinali
--
It is upon moral qualities that a society is ultimately founded. All
the prosperity and technological sophistication in the world is of no
use without that foundation.
-- Miss Matheson, The Diamond Age, Neil Stephenson
<circuit.png><currents.png>_______________________________________________
time-nuts mailing list -- time-nuts@febo.com
To unsubscribe, go to https://www.febo.com/cgi-bin/mailman/listinfo/time-nuts
and follow the instructions there.
On Fri, 1 Jul 2016 12:28:44 -0400
Bob Camp kb8tq@n1k.org wrote:
Real cascode circuits can be built with RF transistors. They also can be simulated.
Simulating them with the “standard” models is a PIA. The issue is that the inductance
of the package is not de-embedded from the test “socket” as carefully as it might be.
There is also the somewhat non-intuitive need to stick a low value resistor in the base.
Done properly, they are very reproducible and reasonably insensitive to load.
Thanks! That resistor in the base did the trick!
Am I right in the assumption that the resistor gives the transistor
some negative feedback and thus prevents it from oscillating?
Attila Kinali
--
Malek's Law:
Any simple idea will be worded in the most complicated way.
Hi
Yes, you can also look at it as “damping” or" de-Q-ing”. You trade off a bit of isolation
for stability. Put another way, the resistor will take the isolation of the stage down a bit.
Another practical point on the stage - you want the base bypass as close to the end of that
resistor as you can get it. You want the resistor right up against the base. Inductance in the
base lead is a really bad thing in this case.
Bob
On Jul 2, 2016, at 1:00 PM, Attila Kinali attila@kinali.ch wrote:
On Fri, 1 Jul 2016 12:28:44 -0400
Bob Camp kb8tq@n1k.org wrote:
Real cascode circuits can be built with RF transistors. They also can be simulated.
Simulating them with the “standard” models is a PIA. The issue is that the inductance
of the package is not de-embedded from the test “socket” as carefully as it might be.
There is also the somewhat non-intuitive need to stick a low value resistor in the base.
Done properly, they are very reproducible and reasonably insensitive to load.
Thanks! That resistor in the base did the trick!
Am I right in the assumption that the resistor gives the transistor
some negative feedback and thus prevents it from oscillating?
Attila Kinali
--
Malek's Law:
Any simple idea will be worded in the most complicated way.
time-nuts mailing list -- time-nuts@febo.com
To unsubscribe, go to https://www.febo.com/cgi-bin/mailman/listinfo/time-nuts
and follow the instructions there.