14
\$\begingroup\$

I need to compare a signal to a constant voltage; the signal ranges from 0 to 30mV, and I require a response time of 50ns at 250µV difference. The signal is a triangle wave with a slew rate in the range of a few mV/µs.

When having a look at the comparators offered by TI, they start at an offset voltage of 750µV, with 10ns comparators starting at 3000µV.

When, however, looking at the list of opamps, those start at 1µV offset voltage, with 100MHz amplifiers starting at 100µV.

It's strongly encouraged to use comparators, not op-amps, for comparing signals, so the only option I'm seeing is to pre-amplify my signal with a precision, high-speed op-amp, then use a comparator. However, this sounds wrong. If this is possible, then why don't chip makers offer this as a monolithic solution?

\$\endgroup\$
  • 1
    \$\begingroup\$ Why don't you use op-amp without feedback AS a comparator? Add reference voltage to one pin, input voltage to another and your output should go rail to rail. That way you can target op-amps with as low offset voltage as you wish \$\endgroup\$ – Artūras Jonkus Jan 23 '17 at 10:15
  • 7
    \$\begingroup\$ I've read quite a few PDFs by analog design experts (e.g. Analog Devices AN 849) on this topic, and, to quote it directly, "However, the best advice on using an op amp as comparator is very simple—don’t!". I don't have decades of experience on the subject, so my first instinct is to follow that advice. \$\endgroup\$ – mic_e Jan 23 '17 at 10:22
  • \$\begingroup\$ On second thought, you may be right. If I add hysteresis myself, none of the points from the Application Note seem to apply. I might need a second chained op amp to achieve the desired rise time, though. \$\endgroup\$ – mic_e Jan 23 '17 at 10:26
  • \$\begingroup\$ Answers below seem to deconstruct my comment well \$\endgroup\$ – Artūras Jonkus Jan 23 '17 at 10:35
  • \$\begingroup\$ Is this a one-off board? If it is, maybe you can use a solution that relies on manually trimming the offset. \$\endgroup\$ – pipe Jan 23 '17 at 11:49
28
\$\begingroup\$

High speed with a small difference is difficult to get.

Note that not only do comparators tend to have higher input offset voltages than opamps, but also much higher effective noise, as to get high speed they are wideband beasts.

Oliver Collins produced a paper a couple of decades ago showing that you get much better results, that is less time jitter, if you precede a fast comparator with one or more low noise, low gain opamp stages, each with single pole filtering on the output, to increase the slew rate stage by stage. For any given input slew rate and final comparator, there is an optimum number of stages, gain profile, and selection of RC time constants.

This means that the initial opamps are not used as comparators, but as slope amplifiers, and consequently they do not need the output slew rate or GBW product that would be required for the final comparator.

An example is shown here, for a two stage slope amplifier. No values are given, as the optimum depends on the input slew rate. However, compared to using the output comparator alone, almost any gain profile would be an improvement. If you used for example a gain of 10, followed by a gain of 100, that would be a very reasonable place to start experimenting.

schematic

simulate this circuit – Schematic created using CircuitLab

Obviously the amplifiers will spend a lot of their time in saturation. The key to sizing the RC filters is to choose a time constant such that the time it takes the amplifier to get from saturated to mid point, at the fastest input slew rate, is doubled by the chosen RC. The time constants obviously decrease along the amplifier chain.

The RCs are shown as real filters after the opamp, not a C placed across the feedback gain resistor. This is because this filter continues the high frequency attenuation of noise at 6dB/octave to arbitrarily high frequencies, whereas a capacitor in the feedback loop stops filtering when the frequency gets to unity gain.

Note that using RC filters increases the absolute time delay between the input crossing the threshold and the output detecting it. If you want to minimise this delay, then the RCs should be omitted. However, the noise filtering afforded by the RCs allows you to get better repeatability of the delay from input to output, which manifests itself as lower jitter.

It's only the input opamp that needs high performance in terms of noise and offset voltage, the specs of all the subsequent amplifiers can be relaxed by its gain. Conversely, the first amplifier does not need as high a high slew rate or GBW as the subsequent amplifiers.

The reason that this structure isn't provided commercially is that the performance is so rarely required, and the optimum number of stages is so dependent on the input slew rate and the specifications required, that the market would be tiny and fragmented, and not worth going after. When you need this performance, it's better to build it from the blocks you can get commerically.

Here's the front of the paper, in IEEE Transactions on Communications, Vol 44, No.5 , May 1996, starting page 601, and a summary table showing what performance you get as you change the number of stages of slope amplification, and the gain distribution of the stages. You'll see from table 3 that for the specific case of wanting 1e6 slope amplification, while the performance does continue to improve above 3 stages, the bulk of the improvement has already occurred with only 3 stages.

enter image description here enter image description here

\$\endgroup\$
9
\$\begingroup\$

Those op-amps with very low offset (such as the TLC2652) have much too low a bandwidth for what you want (about 2 MHz) so, realistically you need to compare apples with apples. Also, not specified in that device's data sheet is how the input offset voltage changes with common-mode input voltage. For a comparator, large common-mode offsets are expected and more often than not, an op-amps's offset voltage is specified in ideal signal conditions.

Another fact is that most comparator circuits use hysteresis and this far outweighs any fabulous figure for offset voltage due to the positive feedback from the output being dependent on supply rails.

And here's the main problem with your comparison.

If you look down the TI list after selecting Vos as the filter parameter, the first op-amp that has a bandwidth of 100 MHz or over is the OPA625. Your expectation of 250 uV producing a full swing in 50 ns means that the AC gain at 100 MHz has to be (say) 5 volts/250 uV = 20,000 or 86 dB. Well, the OPA625 has an open loop gain below 0 dB at 100 MHz.

This means that your comparison is again flawed. You need to be realistic when making comparisons. A 100 MHz op-amp is decades inferior to a comparator that can switch its output in 50 ns with a differential input voltage change of 250 uV.

\$\endgroup\$
  • \$\begingroup\$ What about the OPA625 (Vos=100µV, GBW=120Mhz)? \$\endgroup\$ – mic_e Jan 23 '17 at 10:41
  • \$\begingroup\$ Look at my edits. The OPA625 is a useless comparison. \$\endgroup\$ – Andy aka Jan 23 '17 at 10:44
  • \$\begingroup\$ Also, switching on a fractional millivolt without hysteresis will often mean that you are comparing noise to noise in a wide band environment... \$\endgroup\$ – rackandboneman Jan 23 '17 at 17:37
  • \$\begingroup\$ I'm planning on using two comparators and an RS flip flop for manual hysteresis. \$\endgroup\$ – mic_e Jan 24 '17 at 13:10
  • \$\begingroup\$ @mic_e a good way of controlling hysteresis levels! \$\endgroup\$ – Andy aka Jan 24 '17 at 13:26
1
\$\begingroup\$

Lets design that circuit. You want 50 nanosecond response, thus 1/50nS or 20MHz bandwidth is our starting BW.

What NOISE FLOOR? For a low rate of occurrence of FALSE TRIGGERS, the noise power needs to be 10dB weaker than the signal noise (produces 0.1% Bit Errors). Our total integrated noise needs to be 250uV / 10dB or 250uV/3.16 or 80 microVolts RMS. In 20MHz BW.

To find the noise density (and thus the allowed Rnoise), we divide 80uV by sqrt(bW) or 80u/sqrt(20,000,000) or 80u/4,500 or 18 nanoVolts/rtHz. With 1Kohm being 4nanoVolts/rtHz, we can use Rnoise values of 20,000 ohms.

I suggest the RCA/Harris CA3011 wideband amplifier with 3 differential gain stages. Datasheet says it will (typically) limit at 600 microvolts input, and that limited/squarewave output is certainly suitable for driving a fast comparator. Datasheet says NoiseFigure is 9dB at 4.5MHz, given a 1:2 input stepup (resonator PI) from 50 Ohms.

Now, about that uncertain offset voltage.....

\$\endgroup\$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.