An high dynamic range QSD front end

IK1ODO Dec. 2006

A few months ago I bought and built a SoftRock. The first results have been very exciting, the performance was very good. After a couple days of listening in 40m I begun questioning about the dynamic range of the thing: it was intermodulating, and a variable attenuator was needed in input to use it in Europe. So I modified the SoftRock opamp stage, since that was the bottleneck, and started thinking about a better implementation of the QSD. I used ideas from several authors, in particular from the AIQSD of Ahti OH2RZ.
I have then bought five different audio boards, and I'm currently using an EMU 1616 (not 1616M) and a 1212m. The latter uses the same AK5394A selected for HPSDR in Janus, and is a very good product.
With the 1212m the measured dynamic range (noise floor to saturation) in 2.4kHz of band is in the order of 125dB. I started thinking that a front-end with more than 120dB of dynamic range, without preamplifiers, attenuators or AGC would have been a very interesting thing. So this project begun; after about three months I have a prototype with -128 dBm of noise floor in 2.4kHz, and -3dBm of saturation level. I think that this could be a good starting point for a complete high performance QSD receiver; perhaps a good base for the Phoenix board in HPSDR.

The main measured characteristics of the current circuit are:

Low gain audio interface setting (+4dBu f.s.)

RF input level for A/D converter saturation: -3dBm
SFDR: 90dB with -10dBm RF input
Noise floor: -123dBm / 2,4kHz
3' order dynamic range: 105 dB
IIP3: +36dBm

High gain audio interface setting (-10dBu f.s.)

RF input level for A/D converter saturation: -17dBm
SFDR: 100dB with -20dBm RF input
Noise floor: -128dBm / 2,4kHz
3' order dynamic range: 102 dB
IIP3: +25dBm

Circuit description
More tests on mixer dynamics

Circuit description

I used a four-phase, single balanced mixer with a 90° conduction angle, as suggested by Dan N7VE in his papers. The 90° conduction angle should be the best compromise for a low loss, according to simulations by Phil N8VB.
The sampler is an FST3125. Having no internal decoding logic it should have the best switching symmetry.
The divider chain is the classical Johnson ring counter, driving a four-NAND gate. The LO signal by now is delivered by a Rohde&Schwarz SMHU58 signal generator.
The LO signal appears at the antenna port, but it's quite attenuated, with a -60 dBm or lower level. LO harmonics are stronger, but may be stopped by the input low-pass filters. I think that the the low LO radiation indicates a good mixer balance. Here is a typical LO radiation spectrum at the RF port:

The RF signal is translated to a 200 Ohm impedance. The input transformer, not critical, is a trifilar winding on a ferrite binocular. The impedance seen by each opamp input is 800 Ohm (four times the mixer inputr impedance), offering an acceptable match with the opamp noise impedance (more later on this).
A diplexer follows the mixer, performing two functions: it offers a constant impedance to the mixer, and so to the RF port, and avoids out of band signals reaching the opamp. I had many discussions on this point, so I want to clarify my considerations.
The input impedance of the classical Tayloe mixer changes with F, and is difficult to predict. Measurements done on the SoftRock show impedances far from 50 Ohm, so the preselection filters have no definite termination impedance. Then, at sampler outputs there are many fast signals and spikes, mainly at LO frequency, with harmonics up to the UHF region. The opamp shows a virtual short circuits to all signals within its loop bandwidth, but the input stage can't handle fast transients without going in a non-linear operation. So, a diplexer is necessary for this purpose; see below for a measurement of switching transients.
Cutoff frequency of the diplexer is around 250kHz, not critical. It should be out of the audio pass band (96kHz) and inside the GB product of the opamp. The impedance of the diplexer is 200 Ohm, since when a switch is open it must see the RF port impedance. The impedance of the RF port is very close to 50 Ohm from 1 to 50 MHz. Have a look at the impedance plot (S11 at RF port):

Post-mixer amplifier

Following a schematic published by Ahti OH2RZ I used a fully balanced configuration, using an OPA1632. That seemed correct to me, having to drive a balanced load from a balanced source. The opamp is powered from +/- 12V to have enough dynamic range. Many opamps show the best IMD and distortion if the output swing is below Vcc/3, and the high-end audio cards accept 4 Vrms as maximum input level (+10dBu plus a margin). An instrumentation amplifier is not very useful, since they have high noise at low gain settings.
The opamp contributes with his noise to the total noise, of course. How much? The noise in an operational amplifier is modelled with a voltage source (En) in parallel to the inputs and a current source (In) in series. Now, En/In=Rn, that is an equivalent noise resistance. The opamp gives the minimum noise contribution when the source resistance is equal to Rn. For the OPA1632 this value is 3.25 kOhm, difficult to obtain in a wideband RF circuit.
The opamp's noise figure would be 0.53dB if closed on 3.25 kOhm; in our case it sees about 1k Ohm,and the figure becomes 0.72dB. Incidentally, using four expensive AD797 would give a 0.4dB noise improvement; not worth that, in my opinion. If somebody is interested I have an Excel spreadsheet to compute the opamp's NF given In, En and the source R.

Sampler noise

The switches of the FST3xxx family (and similar switches) in theory are not noisy devices. Internally there is a simple CMOS switch connecting input and output, so the only noise produced should be the thermal noise of Ron (4 Ohm), very small. However things are different, as any experimenter that did noise measurements on a QSD receiver knows. The device is noisy, and noise increases with F; the SoftRock had a great success mainly as a 40m receiver.

I think that the noise derives from the switching process itself. Transients are coupled to the channels via the gate capacitance, driven by a 0-5V pulse. The gate-channel C is unknown, but probably is in the 5pF range.

So I started to work on the switching noise, and I discovered that it decreases:
- using 74AC series gates (I started with 74F logic)
- putting 100 Ohm resistors in series to the outputs of the 74AC00
- powering the FST3125 at 6.8V (7V is the maximum allowed)

But, even more interesting, the noise decreases by several dB changing the polarization voltage at the input of the FST3125, traditionally put at Vcc/2. If you look at my circuit there is a potentiometer to vary the Vref. I consider this one a very important point: Vref must be changed increasing the LO frequency. There is a minimum noise value, very broad at 7 MHz, but more critical at 29 MHz. Changing this bias I obtain -127dBm noise floor (in 2.4kHz IF band) from 1.8 to 21MHz, and -126dBm at 29 MHz. The mixer still works at 40MHz, then the flip-flops stop working, so I have no data for 50MHz.

I have no idea of the reason for this behaviour, and about the dependency from the LO frequency. In any case, here are the transients measured on pin 3 of the FST3125, in two different time scales. My 'scope has only a 1GHz band, and I suspect that the band of the transients may be wider. Compare this with the necessity for a diplexer in front of the OPA1632; the transients are in the order of 600mV peak-to-peak.

The same signal at 1ns/div. The slew rate is limited by the 300ps rise time of the oscilloscope.

A rule-of-thumb noise extimate...
If the noise floor is -126dBm in 2.4kHz then the NF is about 14dB. Three dBs are lost in the diplexer resistors, about one dB is the opamp noise contribution, something (another dB?) is contributed by the audio board and from input transformer losses, another dB is lost in the 90° sampling process. My extimate of the mixer ENR is about 8dB; so there is still room for improvement, and indeed tweaking with parameters on a slightly different circuit I arrived at -132 dBm at 14MHz (always in 2.4kHz).

Work to do
The current circuit is mounted dead-bug on a piece of PCB, and it works well. There are still several points for experimenting:
1)To test different bus switches, and different brands of FST3125, or a FST3126 changing the NAND gates to AND. An interesting switch is the FSAV332, but it is small, and requires a PCB. A friend is designing the PCB, so I will do this test soon.
2)Between the NAND and the switch, optimize the value of the dumping resistors. Possibly the slew rate is important.
3)Divider chain. To arrive at 50 MHz requires faster logic, but able to drive the switch.
4)Input filters. I observed a large rise in noise (approx 4dB) fitting a 5-pole low-pass filter. This may be due to the improper termination of switching transients. I hope to find a way to improve the input matching, without having to fit a diplexer for every different filter...

I expect to do more tests in next weeks, and then publish the results.

More tests on mixer dynamics

73 de Marco IK1ODO