When the mask slips - a cautionary (if ancient) tale of IC quality

When an IC looks like a duck but the quack isn't quite right

So let me take you back, back, back to the early days of Digital Audio, circa 1983. The DSP-1 was Neve’s entry into the rapidly-growing market for digital studio equipment. Unlike conventional analog mixing desks, all the processing and routing happened in the digital domain. This, of course, required that analog signals from microphones, direct injection and general old-school musicians’ equipment needed to be converted to a suitable digital format. After all the processing, the results (unless perhaps going to digital tape or a CD mastering box) need to be converted back to analog, for conventional replay equipment, LP cutting lathes and so on.

Back in the day, a lot of circuit boards stuffed with electronics were required. The desk itself was really just control surface stuff. The signal processing sat in racks; the analog front ends in a separate box that could be isolated from the potential interference of so much digital electronics.

The analog-to-digital and digital-to-analog converters themselves were modular products from small-scale start-up suppliers whose names I forget. This was the Wild West of performance digital audio, you couldn’t just go to DigiKey and buy fantastically performing converters for la handful of dollars, as you can today. I dimly recall 16-bit DACs constructed from upper and lower 8-bit converters with linearity correction coefficients stored in UV reprogrammable EPROMs. One thing I do clearly remember is that there was no oversampling of any kind involved in either direction of conversion. The ADCs and the DACs both operated at the native sample rate of 48 ksps. And, therefore there was a really pressing need, in front of the ADCs and following the DACs, for…

Filters!

Which of course is where I got involved, as I worked at the time for a leading specialist supplier of filters of many types, including active.

The specifications required of these lowpass filters, which plugged into the Neve circuit boards were demanding, as you might expect. The anti-aliasing filters needed to have a flat frequency response from DC up to 20 kHz to within a few tenths of a dB (they ended up having an adjustment trimpot accessible from underneath, to ensure consistent response). The stopband requirement was 72 dB of rejection at 24 kHz and above, as you’d expect given their role as anti-aliasing filters. There was also a phase linearity requirement of something like ±1 degree up to 15 kHz, which required an additional allpass block to bring the overall phase shift into line.

Topology-wise, the lowpass section was implemented as a FDNR ladder, 9th order elliptic, using vanilla FET opamps in the core. The carefully-optimized circuit offered sufficiently good dynamic range and large signal linearity at modest current consumption. One we got the encapsulation methodology sorted out, this filter didn’t really cause any issues through the lifetime of the project.

The anti-imaging filter, the output filter for the DAC, that was another matter.

I just stacked up a segment of an article explaining why the frequency response of a sample-and-held DAC droops down at high frequencies. This digital audio system with its 48 ksps DACs was no different. Measured with no output filter, the DAC frequency response at the 20 kHz upper specification limit was around -4 dB, which was completely unacceptable for a piece of state-of-the-art professional audio gear. So the specification for this anti-imaging filter was written to ensure that the overall response, when driven by the DAC used, would meet the same kind of passband flatness specification (and phase linearity) as the companion anti-aliasing filter. The one relaxation was that the -72 dB stopband began at 28 kHz - the frequency at which a 20 kHz audio signal would image to. The noise and linearity performance also needed, of course, to be excellent.

The design methodology needed to be a little different for this filter. Adding a response ‘boosting’ circuit to cancel out the droop was problematic; put it at the input and high frequency overload is a problem. Put it at the output and a significant degradation of noise would result. And a circuit to match the required inverse of the practical droop curve with the necessary closeness of fit was also not straightforward.

I decided to implement this filter as a cascade of 2nd-order filter sections (‘biquads’), a very conventional approach that you’ll see deployed on practically every online filter design tool out there. I wrote the necessary optimization routine to ‘bend’ the coefficients of a fairly standard 8th order elliptic filter so that the frequency response had the right inverse response for the droop of the DAC back-end.

The signal handling performance of such a filter cascade is strongly dependent on the way you deploy the individual poles and zeros of the transfer function along the biquads in the cascade, so that needed to be done carefully. The intrinsic sensitivity and noise characteristics of various biquad topologies vary widely; several different biquad types were used.

And now I finally get to the point of all this. To meet the audio specifications if this filter, where all the opamps are participating in a more direct way than happens in the FDNR design, I needed something rather ‘better’ than the vanilla FET devices used in the anti-aliasing filter. Step forward, your friend and mine, the…

NE5532

This workhorse dual opamp of audio electronics was a lot newer in those days than in it now, and had no real competition (perhaps Nat Semi’s LM833). The original Philips TDA1034 became the NE5534 when their US partners, Signetics, shoehorned it into the product range. The NE5532 was the dual version, with enough interna compensation to keep it happy at unity gain (it has a slightly awkward second pole that caused me a headache at one point, let’s leave that for another day).

So, we build prototypes and early batches with Signetics NE5532s, and very good the performance was too. Until, after one delivery…

The Phone Call

where my contact at Neve engineering explained that the THD they were seeing on their more detailed goods-inward testing had increased by something like a factor of 8, from (if I remember, 40 years is a long time ago) around 0.005% to 0.04%. I did a detailed review of the latest batch against me reference samples, and sure enough, the distortion was too high.

Of course, these were the wonder days of DIP packaged devices. I already had socketed prototype boards so we started looking at ICs. We found that Signetics NE5532s with a few newer date codes (all the ones in sock at distys at the time) systematically delivered poorer performance than older parts.

We needed a knight in shining armor to ride to the rescue; on cue…

Texas Instruments

by that time were producing their own NE5532, so popular had this device become. So we got some of them and phew! All was well. So we started using those, and Neve were happy once more.

I was just a lowly engineer then so I didn’t get involved in any shenanigans with the supplier except to produce the incontrovertible proof that the part had changed in an unacceptable way. I think we got some recompense for the out of spec batch that we had built.

Some time after that, we heard that the ‘reason’ was that the fab producing the Signetics NE5532s had ‘lost’ parts of the mask set and/or diffusion information for the parts, and someone had decided to recreate it without creating any kind of trail or errata for the part. I think they decided that this was not in itself a process failure (unlike a sodium contamination / migration issue on some Nat Semi parts we’d experienced a few years before, might have been LM308s dying in the field after having power applied for a while, don’t quote me) and wasn’t actually a quality issue, they hadn’t noticed, and the parts they made had not actually failed any production tests.

If there’s a moral to this ancient tale, I suppose it’s “don’t make assumptions”, especially about distortion. We often extrapolate from multiple pieces of knowledge in order to determine what we think the distortion performance of a circuit will be, given a few specification points and an approximate schematic of the part. Sadly, most vendor SPICE models will not give you enough insight. That conundrum is enough for an entire book (though there have been some good pieces on noise modeling of the 5532 in AudioXpress magazine over the past few years).

Thanks for reading! — K