Recently, a VDSL line driver design of mine underwent reliability testing in the burn-in oven. What we do is load a large board with 100 parts and make them follow a dynamic signal at high temperature for 1,000 hours, expecting no failures.
At the first 156-hour pull we had failures. Subsequent pulls yielded more failures. The process we use for VDSL is a rugged one, and offers 7GHz complementary bipolar devices. When designing, we harden the IC to tolerate routinely large output currents, ESD, and lightning strokes, so burn-in failures are a surprise.
When re-loading the burn-in board for a second go, our product engineer noticed that N devices did not draw N times the expected supply current, but more than that. Starting back with one loaded part, he got the expected current. Two devices, twice. But three devices, an excess of supply current. So he came to me with this information, expecting answers.
I remembered I had an odd experience with the part some time ago in its first days of evaluation. I had the part in a test board and was measuring its distortion, and it was 20dB higher than expected. This was surprising, because nowadays the device modeling and simulation accuracies are pretty good. Further, the quiescent supply current was about 30 percent high. We nail supply current to within 5 percent in the design process. This is a clue that the circuit is oscillating, but my 400MHz oscilloscope (with a good 50Ω interconnect system) showed none.
When looking for oscillations I have always used my fingers to touch and thereby capacitively load various parts of the board connections. If you're looking for oscillations, you should be able to create one or perhaps suppress it. The circuit had a nominal bandwidth of 150MHz, so I figured my 'scope and my fingering should deliver an oscillation. I could create and observe an oscillation by loading the negative feedback terminal at the roughly expected 100MHz, but without stimulus there was none.
But when I removed my finger the supply current had reduced. After a bit more probing, the original supply current had returned. I found I could induce the higher supply current by touching here, then return it to a lower supply current by touching there. However, I still could not see oscillation on the 'scope. Remembering that the IC was made of 7GHz devices, I guessed that I needed a wider-band measurement system to find oscillation.
The best way to find an oscillation is to use a spectrum analyzer, which can sweep vast amounts of spectrum and has a very wide amplitude range. However, the analyzer has a 50Ω input impedance, and randomly connecting that low impedance to various parts of a circuit can suppress the oscillation you're looking for. So I made a "sniffer," a coaxial cable whose end is cut off and the inside wire is made into a loop and shorted to the shield. This is a single-turn transformer secondary that can convert AC magnetic fields to voltage displayable by the spectrum analyzer, without appreciably loading the circuit. You wave the loop over the suspected oscillation source and look for a spike on the spectrum. You have to discount a lot of radio signals, but I found a strong 1.8GHz spike. To test its source I turned off the power to the test board and the signal disappeared. Power on, signal back. Oscillation found. I would never have seen this on the oscilloscope.
I also found I could change the frequency of the oscillation by touching the circuit in the previous place that caused the supply current to shift, and amazingly the oscillation frequency dipped to 1.3GHz. Interesting, there were two different oscillation modes that were bistable.
It turns out this 150MHz circuit was not oscillating through feedback, but it was only the input pin that oscillated. I could stop the oscillation by touching the input line, although remote parts of the line did not suppress oscillation when touched. It appears the transistors on an IC connected to an input or output can present negative-real impedances to the outside world at some ranges of very high frequency, and external resonances at odd frequencies can conspire with this negative impedance to form a one-pin oscillator. Current-feedback amplifiers have this possibility at their positive inputs, at frequencies around ft/4. I added a moderate-value series resistor to the input on the chip in a re-design. Problem solved.
So back to the present and the burn-in issue. The product engineer made a sniffer to look for oscillations. He does find them, but at 150MHz, and only when three or more parts are loaded. Moving the sniffer shows that the region around the parts and the supply line are all hot with signal.
The board has four layers of copper: one is the ground plane. Another is the V+ supply plane. The top is general interconnect, and the bottom is interconnect and V- supply interconnect. The inductance of the planes is probably very low and won't contribute much inductance between adjacent devices, but the V- interconnect is inductive between devices. The general layout and choice of high-temperature components and sockets is not what I'd consider low-inductance at 150MHz; even if one IC does not oscillate in the board, clearly many devices are having a party and conducting a gang oscillation. Not too surprising; the amplifiers have lots of gain at 150MHz but very little supply rejection at this frequency. Oscillation energy must be sloshing back and forth across the board with each amplifier adding its bit of gyration to the wave.
So how to fix? In high-gain amplifier chains, the supplies are routinely isolated from amplifiers with a bypass capacitor at the amplifier (to keep supply AC currents circulating locally to ground) and a series resistor or lossy choke from amplifier supply to supply bus (to keep hostile local noise from the bus and attenuate bus noise coming in). Our board is not easily modified. Inserting things in series is difficult when modifying a board. How about parallel? We have this high-Q supply interconnect network -- maybe we could lower the Q and make it lossy at 150MHz. We put snubbers at every bypass capacitor and the oscillation ceased.
A snubber is a low-value resistor (20 to 50Ω) in series with a moderately large capacitor (500pF in this example), all put in parallel with a node you want de-Q'ed. The resistor is lossy and damps resonances, and the capacitor prevents DC or low-frequency loading.
I've never dealt with a multi-amplifier supply oscillation before, so this was interesting. Usually amplifiers oscillate alone, rarely with another. Guess there is a first time for everything. Please let me know if you've any problems with spurious oscillation like this.