dCS 900B
A single-purpose box that did one thing - the highest-quality analogue-to-digital conversion its designers could manage - and did it well enough to earn a following among the engineers whose recordings would expose anything less.
Overview
The 900B is a stereo analogue-to-digital converter from Data Conversion Systems. It sits in a black single-unit-high rackmounting enclosure that looks like a lot of box for something that is "only an A-D convertor." Inside is a dense field of electronics. dCS left it fanless and free of ventilation slots, so there is no noise and no internal dust to worry about. Construction is sturdy, though the period reviewer who bench-tested it recommended rear support for rackmounting given the weight of the thing.
The architecture is the story. The 900B is a 128-times oversampling, noise-shaping converter. It samples the input 128 times faster than the designated sampling rate, then decimates back down. The reason is the anti-aliasing filter. A conventional converter needs a very sharp analogue filter just above the audio band to keep out-of-band signals from folding back into the audible range, and those sharp filters introduce group delay that smears the phase relationship between, say, a 1 kHz fundamental and a 20 kHz harmonic arriving in phase. By oversampling to 128 times 48 kHz - 6.144 MHz - the input filter can be simple, cheap, and stable, with an almost negligible slope and therefore almost no delay. The decimation that follows is done with a nearly ideal digital filter whose phase shift is close to linear, so the frequencies that pass through stay in time with each other. According to the manufacturer's specification, group delay is under 1 ms, and the review measurements confirmed both the delay figure and the phase linearity.
The measured performance is serious. Signal-to-noise ratio is quoted at 105 dB over a 0 to 20 kHz bandwidth. Spurious levels sit at −100 dBFS at full-scale input, dropping to −125 dBFS and below as the input level falls away - typically −130 dBFS. Common-mode rejection exceeds 90 dB at 50 Hz. Frequency response is flat to within −1 dB from 1 Hz to 20 kHz at 48 kHz and 47.952 kHz sampling, to 18 kHz at 44.1 and 44.056 kHz, and to 13.33 kHz at 32 kHz. The reviewer's own linearity testing turned up a curiosity: the two channels did not behave identically. Reducing output word length from 24 bits to 16 bits caused channel 1 to lose linearity about 10 dB earlier than before, while channel 2 was unaffected. The review asked why the channels should differ so much in low-level linearity, and which behaviour was the norm.
Operation is minimal. There are only five user controls, and all of them live on the rear panel: a sampling-rate selector (48 kHz or 44.1 kHz), a word-length switch (16, 18, or 24 bits) that alters the noise shaping and dither to suit the target word length, and separate multi-turn screwdriver presets for channel 1 and channel 2 sensitivity. Sensitivity is adjustable from +14 dBu to +26 dBu for full-scale modulation. A switch underneath selects the upper or lower of two mains voltages, and power enters through a filtered IEC socket with its own switch and fuse assembly. Alternative sampling rates could be supplied, but only two could be fitted within a given unit. Once the box is rackmounted, none of these controls is easily reachable. The front panel is blank apart from a few LEDs for sampling rate, master/slave status, and overload.
The overload behavior was clearly thought through. On the main unit, overload indication is momentary and held for about a second. A rear-panel Remote connector accepts an optional small remote box carrying two overload LEDs, a power indicator, and a reset switch on top. On that remote, the overload indicators latch, and a poke of the reset button clears them. dCS state that overload detection is very comprehensive, and the reviewer's experience confirmed it: across various waveforms, a headroom of from 0.1 dB to 2 dB was necessary between nominal output levels and the detection of overload relative to a single sine-wave input.
The analogue inputs are XLR-3 female connectors (pin 2 hot, differential input impedance above 10 kΩ, common-mode above 2 kΩ, damage level above 46 dBu). Digital outputs are AES/EBU on a 110 Ω line, SDIF-2 with a word-clock on BNC, and SPDIF on a 75 Ω RCA phono. Word-clock in and out are on BNC, with time skew under 100 ns referenced to word-clock in/out. Sampling-frequency accuracy is ±10 ppm in master mode as shipped, holding to ±10 ppm per year and ±15 ppm over temperature. In slave mode the unit locks within ±300 ppm of nominal in under 30 seconds. Dimensions are 44 × 430 × 390 mm (H × W × D), and the unit is 1U 19-inch rackmountable. Mains versions were offered at 120/240 V (UK), 100/220 V (Japan), and 110/230 V elsewhere. The case is simply made of steel with an aluminum front panel, held together with plenty of screws, and internally there are three large PCBs.
Studio Sound (April 1993, p.61) put the 900B on the bench and found that it accomplishes its goal. Every aspect named in the manual as a design target was, in the reviewer's assessment, implemented better than other similar products tested to that date. The same review noted the difficulty of devising a sensible listening test, since the 900B has only analogue inputs and digital outputs and is therefore at the mercy of the D-A converters used for both source material and monitoring. The reviewer declined to pronounce on sound quality and instead pointed to the box's following among classical recording engineers, the very people whose programme material would reveal the deficiencies dCS set out to overcome. The price was £2,600 in the UK.
In a letter to the same publication (Studio Sound, May 1993, p.54), Paul Maddox of Data Conversion Systems noted that the 900B's published specification was based on tests carried out with the company's own dCS 250 Converter Test Systems, and that review results found to be "better than specified" were attributed to continued product development. He also confirmed that the 900B's group delay had been measured at 570 ±5 µs at the 48 kHz sample frequency, comfortably inside the under-1 ms figure. His summary of the review's significance: simple operation and reliability delivering long-term financial savings, and the argument that all recordings contribute to history, so nothing but the best would do.
The 900B did not sit still. A retrofittable digital filter board was introduced that added second- and third-order noise-shaping algorithms to the original first-order design, plus a ninth-order psychoacoustically shaped output intended for end-user applications such as CD premastering. New units supplied with the board factory-fitted were designated the dCS 900C (Studio Sound, November 1994, p.34). The same product guide lists the dCS 902, a double-sampling-frequency version of the 900B built for use with "high speed" digital recorders. A later Studio Sound notice (October 1993, p.16) describes the 902 as based on the 900B architecture, using a discrete dCS proprietary oversampling converter, with 88.2 and 96 kHz sample frequencies, master or slave operation, selectable output word width, remote overload monitoring, and double-speed AES/EBU, SPDIF, and SDIF-2 interfaces. It also avoids gain-ranging techniques to prevent noise pumping.
Where other products tried to accomplish several tasks in one box, the 900B went for high audio performance and hit it. The reviewer's closing question was whether a recording engineer's customers would notice the difference a £2,600 A-D converter makes. The review's own answer is a qualified yes: for the esoteric hi-fi crowd, certainly; for potentially historic recordings, anything but the best would not do; and for the rest, perhaps the sensible move was to wait for semiconductor technology to catch up with the dCS crowd.
eBay Listings
As an eBay Partner, we earn from qualifying purchases. This helps support our independent vintage technology research.