It wasn't that long ago that a 16-bit monolithic digital/analog converter (DAC) was an impressive achievement. Please understand that I'm talking about "true" 16-bit performance, with 1 LSB integral and differential nonlinearity (INL and DNL), not just "apparent" 16-bit resolution.
But that's old news. We now have true 18-bit monolithic DACs from two leading vendors, Analog Devices and Linear Technology Corp. (and even 20 bits, depending how you use the bits). Their respective AD5791 and LTC2757 DACs achieve this impressive performance, and at what seem to me to be modest prices. Applications include precision test, measurement, control, and instrumentation situations.
When you really dig into some of today's critical projects, the x4 increase in resolution you get by leaping from 16 to 18 bits is needed and worthwhile–if you can actually take advantage of it by proper circuit design and layout. Keep in mind that 18 bits is one part in 262,144, or 4 parts per million (ppm), or 0.0004%.
All of which brings us to the eternal question that plagues and challenges developers of any high-precision, high-accuracy instrumentation: how do you know it's really delivering that performance? I got a sense of the challenge in a conversion with analog circuit expert (many say "genius") Jim Williams of LTC. He described how he developed a system to measure the performance of the 18-bit DAC. What he was looking to do was to measure settling time to 20-bit (1 ppm) resolution for settling times as short as 265 nsec.
Jim said he actually worked on the ideas and design for several years; some of his first-pass circuit boards are between five and ten years old. You can read all about his efforts and techniques in the detailed 36-page report, Application Note 120, "1ppm Settling Time Measurement for a Monolithic 18-Bit DAC: When Does the Last Angel Stop Dancing on a Speeding Pinhead?"
It's that sub-head that really says it, added Jim. When you are in the 18/20-bit world, it's hard to know what's real and what's not. Every perturbation, vibration, casual motion, thermal variation, and other normally ignorable flutter can affect those last bits in nearly impossible-to-determine or perceive ways. At some point, you're moved beyond the certainty of technology and into the realm of faith and levels of confidence.
In short, it's the eternal test-and-measurement question: how do you test the system? And how do you test the tester? Or, as asked by the Roman poet Juvenal about 2000 years ago, "Quis custodiet ipsos custodes?" "Who watches the watchmen?" "Who guards the guardians?"♦