High Voltage Signal Source
When it came time to test the lab prototype described in part 1, I realized that I had no signal generator with enough output voltage swing of any waveform to exercise the circuit. I have generators that produce various waveforms to ±10 VP-P. It is time to come up with an amplifier that can cleanly reproduce waveforms at large amplitudes. Figure 4 shows a high voltage discrete realization of a current-feedback amplifier (CFA).
CFAs have fabulously high slew rate and, usually, wide bandwidth.1 Because we are using high voltage transistors, though, the bandwidth is modest. High voltage transistors have higher parasitic capacitances and lower Fts than lower voltage types.
Some warnings here. There is no current or dissipation limitation built into the circuit, so heavy sustained load currents more than 10 mA will burn out the output stage and maybe more stages. Further, it’s best to not add bypass capacitors less than 0.1 µF to the high-voltage supplies. A short-circuit can cause welding if a big capacitor is used. Unfortunately, I had to add 100 µF bypass capacitors to the high voltage supplies to suppress second- harmonic distortion so watch out for this. I crank the lab supplies up and down by hand to avoid hard turn-ons and turn-offs. Please note that even 50 V can cause enough current through a human to cause heart arrest. It’s best to turn the current limit of the high voltage supply down to 60 mA as well. 50 V is high enough to be respected.
In Fig. 4, the ADA4898 op amp controls the current-feedback amplifier (CFA) and keeps its accuracies and distortions controlled. CFAs generally have high dc errors and poor settling to high accuracies; the op amp fixes those.
The positive input of the CFA is node n25 and its negative input is n5 (yes, that’s an input). By itself, RFF and RGG set the gain of the internal CFA to about 27. This high gain allows the controlling op amp’s output to swing only ±2 V. The CFA could have been set to higher gain to unburden the control amplifier further, but then the CFA would lose bandwidth and increase its distortion. Overall gain is set by RF and RG and is 20. CTWEAK and CTWEAK2 work with RF to remove the phase lag of the CFA from the overall op amp feedback above 215 kHz, enhancing op amp stability.
The n13 is the CFA gain node and is driven by current mirrors involving Q1/Q2/Q20 and Q11/Q12/Q19.
Q7/Q8/Q10/Q13 form the output buffer as a compound complementary emitter follower. There is no current limit circuitry—don’t short the output to anything.
The CFA section of the high-voltage amplifier has a 35 MHz, –3 dB bandwidth and does not peak, on its own. The overall circuit has a 33 MHz, –3 dB bandwidth but with 8 dB of peaking. Normally, the second amplifier of a composite amplifier design has at least 3× the bandwidth of the input control amplifier to avoid the peaking; but we could not get so favorable of a ratio. At least the 8 dB peak does not have a high Q and ringing damps reasonably fast. The intended 100 kHz signals are reproduced just fine below the peaking frequency. The distortion at an output of 80 V p-p at 100 kHz measured –82 dBc, dropping to –100 dBc for outputs of 32 VP-P and less at 100 kHz. A square-wave response has a ~60% overshoot for fast edges and little to no overshoot occurs with output slew rates less than 250 V/µs. The maximum slew rate is about 1900 V/µs.
Now that we have big signals, how do we use ordinary lab gear to measure the ±40 V outputs? Neither the high voltage amplifier nor the high voltage buffer should output more than 10 mA, nor can they work into more than ~40 pF load stably. At 27 pF/ft, coaxial cables are too capacitive. An oscilloscope ÷10 probe will have only ~15 pF||10 MΩ loading, so that will be fine for coupling to an oscilloscope.
For measuring distortion, none of the audio analyzers in our lab can beat –80 dBc at 100 kHz, so we must turn to spectrum analyzers. These unfortunately have only 50 Ω inputs—far too low for our circuits to drive. My solution was to raise the impedance to 5050 Ω (see Figure 5); that is, place a 5 kΩ divider resistor between the signal and the 50 Ω analyzer input, making close to a ÷100 divider. It is important that the 5 kΩ resistor not exhibit thermal shifts during low frequency signals, because these shifts are VOUT2² related and cause even harmonics. I chose to put five 1 kΩ, 2 W resistors in series to make RDIVIDER. A 2 W resistor will have about 37°C/W thermal resistance, and the five 1 kΩ resistors have 7.5°C/W thermal resistance. With a ±40 V sine wave across it, there is a 160 mW dissipation, and that will cause 7.5 × 0.16 = 1.2°C heating in the resistors. They have around 100 ppm/°C resistance shift, so at dc there would be a 120 ppm shift, or around 0.01% nonlinearity and –80 dBc generated distortion. How can this ever be accurate enough for our measurements? The good news is that the divider resistors have fairly long thermal time constants, and we expect little actual resistor shift in the middle of 100 kHz cycles. We would ironically see worse distortion at lower frequencies, probably 1 kHz and below.
The 80 VP-P signal had to be attenuated anyway because of limited analyzer input range, but it is still too large to get the best spectrum analyzer performance. Our analyzer can only offer –80 dBc distortion unaided, as a trade-off between its noise swamping the harmonics and large inputs causing additional distortion. A solution is to place a 100 kHz trap at the analyzer input to kill the fundamental amplitude. With less than a few millivolts of signal (harmonics only) we can approach –120 dBc measurement range. Figure 5 shows the test setup.
Figure 5. Adding the high-voltage amplifier, a bootstrapped buffer, and passive filters provided the needed signal path to perform a distortion test
The generator drives RTERM through a low-pass filter LINPUT and CINPUT that attenuates our generator’s harmonics of 100 kHz. This improves its distortion to about –113 dBc, lower than the circuits to be measured. The cleaned-up signal is boosted by the high voltage amplifier and passed by the buffer, which drives the divider.
The inductors are constructed of magnet wire wound on large bobbins intended for power E-I cores. Core materials of any kind cannot be used due to added distortions; air-wound is mandatory. You just wind and measure repeatedly.
LTRAP was found to be magnetically radiating harmonics to adjacent, sloppy unshielded wiring, my usual approach, so I put the trap components in a cookie tin with a grounded BNC jack connection. We use cookie tins in our lab; I like roasting pots, but any shielding steel box would do.
For calibration, I replaced the bootstrapped amplifiers with a through wire and logged the gain from the voltage at RTERM to the spectrum analyzer inputs at second through fourth harmonic frequencies. When measuring a harmonic in the distortion test, I use the stored gain at that frequency to infer the harmonic content at the output of the buffer. I have an oscilloscope monitoring the amplitude of the buffer fundamental frequency output, rms the normalized harmonics, and divide by fundamental amplitude to get overall distortion.
Using the setup of Fig. 5, the spectrum analyzer showed a distortion of –81 dBc at 70 V p-p and 80 V p-p output, –82 dBc at 50 V p-p and 60 V p-p out, and –86.5 dBc for 16 V p-p and 32 V p-p out, all at 100 kHz.
DC linearity, gain accuracy, and input range were then measured. Figure 6 shows the input offset of the buffer as we sweep the input dc signal.
Multimeters can’t easily resolve sub-microvolt variations against a background of ±40 V signals, but since this is a buffer we can simply connect a voltmeter from input to output to find offset shifts and use a sensitive range. The common-mode rejection of my multimeter was less than 1 µV for ±40 V inputs (inputs shorted for that test).
The perturbations in the curve are caused by low frequency noise and especially thermal perturbations. Just having a human in proximity or air conditioning can cause drafts and thermal variations that cause Seebeck and thermocouple voltage errors in a circuit at the microvolt level. I did not have a good shield or screen room, but I did cover my circuits with some clothing to prevent drafts. Even still, there is 0.6 µV rms wander in the results.
Amidst the noise, the unloaded (green) curve suggests a gain error of ~0.03 ppm. Not bad. The un-bootstrapped LTC6240 would have a nominal gain error of 5.6 ppm, and worst-case 100 ppm due to CMRR error. When loaded with 50 kΩ (purple), we see a gain error of –0.38 ppm. The loaded gain error is equivalent to an output impedance of 0.02 Ω. It’s hard to know what the source of that 0.02 Ω comes from—it could be load currents modulating VP or VM and acting through common-mode rejection or gain limitation processes within the LTC6240, or it could simply be wire and circuit board resistances. In any event, to keep gain precise we could connect the feedback of the LTC6240 remotely to the final load to effect a Kelvin connection.
Figure 7 shows the small-signal pulse response.
All apologies for the ringing of the green channel, which is the output of the high voltage amplifier. It doesn’t ring on its own, I just had a mediocre oscilloscope probe and board-to-board grounding. The yellow channel is the buffer output, and it’s a simple exponential dominated by the CIN + RIN.
Figure 8 shows the large signal pulse response with an input slew rate of ±32 V/µs—a nice, smooth response.
Figure 9 shows the buffer response to an overloading slew rate. An 80 V p-p output at 100 kHz demands a peak slew rate of ±25 V/µs, within the ±32 V/µs capability shown.
Note that the input filter limits the overloading slew rate to an amount the buffer can follow. The ripples are artifacts of the inability of the bootstrap circuitry to follow the output slew, which causes input headroom overloads repeatedly during the slew. Reducing CIN forces more input slew rate and the bootstrap circuitry will not follow, causing much uglier rippling.
A method to effectively bootstrap a low voltage op amp buffer to become a high voltage buffer has been shown. We have taken an op amp with rare input characteristics and elevated it to have higher voltage range, better gain accuracy, higher slew rate, and less distortion than the original op amp.
- Barry Harvey. “Application Note AN1106: Practical Current Feedback Amplifier Design Considerations.” Renesas, March 24,
About the Author
Barry Harvey has worked as an analog IC designer, designing high speed op amps, voltage references, mixed-signal circuits, video circuits, DSL line drivers, DACs, sample-and-hold amplifiers, multipliers, and more. He has an MSEE from Stanford University. He holds more than 20 patents and has published about as many articles and papers. Barry’s hobbies include repairing used test equipment, playing guitar, and working on Arduino-related projects. He can be reached at email@example.com.