NOS multibit DACs explained: TDA1541, PCM1704 and the classic chips

Non-oversampling (NOS) and multibit converters are popular again, and they bring a vocabulary that is easy to confuse: NOS, multibit, R-2R, ladder, delta-sigma, bitstream. This guide sorts out the terms, describes the classic Philips and Burr-Brown chips that NOS designs are built around, and sets out plainly what a NOS converter gives and what it costs.

What does NOS (non-oversampling) mean?

A CD stores 44,100 samples per second for each channel. Almost every modern DAC first passes those samples through a digital interpolation filter. The filter calculates extra samples between the real ones (oversampling, typically by eight times or more) and removes the mirror images of the audio band that sampling creates. Only then does the converter turn numbers into a current or a voltage.

A non-oversampling DAC leaves that stage out. Each original sample is converted as it arrives and held until the next one comes. Engineers call this a zero-order hold: the waveform is drawn as a fine staircase. Any smoothing happens afterwards in the analogue domain, in the output stage, in a gentle analogue filter, or simply in the limited bandwidth of the rest of the system.

In this article NOS means non-oversampling. It has nothing to do with “new old stock”, the term for unused vintage tubes, although NOS converters often use those as well.

Multibit, R-2R and delta-sigma: what is the difference?

The question “R-2R or multibit?” mixes two levels. Multibit describes what the converter does: it resolves each sample in one step, with one weighted element per bit, so a 16-bit word switches sixteen weighted currents at once. R-2R describes one way of making those weights: a ladder of resistors of value R and 2R. Every R-2R DAC is therefore multibit, but not every multibit DAC is a resistor ladder.

• Multibit chips with weighted current sources. The Philips TDA1540 and TDA1541 generate binary-weighted currents and keep them accurate with dynamic element matching, which continuously swaps the most significant current sources so that their small errors average out.
• Multibit chips with laser-trimmed ladders. The Burr-Brown PCM63 and PCM1704 contain resistor networks trimmed at the factory, with extra circuitry around the zero crossing, where a ladder’s errors matter most for quiet music.
• Discrete R-2R. A ladder built on the circuit board from many precision resistors and switched by logic. Most current products sold as “R2R DACs” work this way.
• Delta-sigma. The converter uses very few levels at a very high rate and pushes the resulting noise above the audio band (noise shaping). Accuracy comes from timing rather than from matched parts. Philips called its 1-bit version “bitstream”. Nearly all modern chips are delta-sigma; some use a few bits rather than one, such as the Cirrus Logic CS4397, a multibit delta-sigma design.

So “1-bit or multibit” is a real difference in architecture, while “R-2R or multibit chip” is mostly a difference in where the ladder lives: on silicon or on a circuit board.

The classic chips at a glance

All five chips are long out of production. Building with them today means sourcing old stock and checking every part, since remarked chips circulate.

Philips TDA1540

The converter of the first Philips CD players. It is a 14-bit chip; in those players it worked with four-times oversampling and noise shaping to reach 16-bit performance. Fed 16-bit words directly, a 14-bit chip cannot resolve the two lowest bits on its own, so how a NOS design handles them is a design decision. Listeners often describe its sound as natural and unforced.

Philips TDA1541 and TDA1541A

The 16-bit stereo successor, and the chip most associated with NOS converters. Selected grades of the TDA1541A, marked S1 and S2 (the “single crown” and “double crown” parts), were sorted for better linearity. It is commonly described as fluid and full-toned, with strong dynamics; some listeners find its treble less extended than that of modern chips.

Philips TDA1547

The TDA1547, often called DAC7, belongs to a different family. It is a 1-bit bitstream converter, not a multibit chip, and it was designed to receive a high-rate, noise-shaped signal from a companion processor such as the SAA7350. Oversampling is therefore part of how it works. Owners of the 1990s players built on it describe a smooth, dark and refined presentation. A comparison of the TDA1541 and the TDA1547 is a comparison of two architectures, not two generations of the same idea.

Burr-Brown PCM63

A 20-bit mono chip with the Colinear design: two complementary ladders arranged so that the major-carry error at the zero crossing, the point every quiet signal passes through, is kept small. Parts were graded by linearity, the K grade being the most selected. It is commonly described as large-scale and weighty. Our notes on the PCM63 datasheet go further into the grades and the current output.

Burr-Brown PCM1704

A 24-bit mono sign-magnitude chip, the last of the classic high-end ladder DACs. It is commonly described as precise, with controlled bass and strong attack. Asked “PCM63 or PCM1704?”, listeners usually set the PCM63’s scale and body against the PCM1704’s control and detail.

What does NOS give you?

• No digital filter in the path. Steep oversampling filters ring before and after a transient. A NOS converter has no such filter, so its impulse response shows no pre-ringing.
• A short, simple signal path. No interpolation chip, no extra clock domain, fewer parts between the data and the output.
• The character of the chip. Many listeners prefer the tonal density of a good multibit chip at low level, and NOS lets it reach the output unchanged.

What does NOS cost?

• Treble roll-off. The zero-order hold has a gentle high-frequency slope of its own. For an ideal converter at 44.1 kHz it amounts to about 0.7 dB at 10 kHz and about 3.2 dB at 20 kHz. This is one reason NOS and some R-2R designs are described as soft rather than sharp in the treble.
• Image frequencies. Without a digital filter, mirror images of the audio band appear above 22.05 kHz. A 20 kHz tone is accompanied by an image at 24.1 kHz at only a slightly lower level. The images are inaudible in themselves, but they reach the amplifier and the tweeter, and any part that is not linear at those frequencies can fold them back into the audio band. Designers either add a gentle analogue filter or accept them for the sake of a simpler path.
• Measured performance. Distortion and noise are higher than in a modern delta-sigma converter, and true 16-bit linearity depends on the chip and its grade.
• Formats. NOS designs are built for 16-bit / 44.1 kHz. High sample rates and DSD call for a different architecture. Our article on SACD and DSD explains why.

Why the output stage matters so much

All five chips deliver a current, not a voltage. The output stage must convert that current to a voltage (I/V conversion), add gain and drive the cable to the amplifier. The conversion can be done by a resistor, an op-amp, a transformer or a tube stage. In a NOS converter the output stage also receives the image frequencies, so its bandwidth and its linearity above 20 kHz matter more than they would after a digital filter.

Tube or solid-state output stage?

A solid-state stage offers lower output impedance and lower noise, and it is indifferent to long cables. A tube stage adds mostly low-order harmonics, which many listeners hear as more body and air around voices; it needs warm-up, and its tubes age. Neither is correct in itself. The choice depends on the amplifier that follows and on how much colour the rest of the system already brings.

NOS, oversampling or programmable logic?

If most of your library is CD and CD-quality streams, a NOS multibit converter plays it without conversion to another rate. If you listen to high-resolution files or DSD, an oversampling design suits you better. Programmable logic (FPGA) converters sit at the other end: the designer writes the filter, and some convert all PCM to DSD. The question “FPGA or R-2R?” is really a question of philosophy: computed filtering against no filtering at all.

Which Demograf Audio piece fits

• The Reference Tube DAC is one chassis built in five versions, each named after its chip: Calliope (TDA1540), Persephone (TDA1541), Aphrodite (TDA1547), Artemis (PCM63) and Katerina (PCM1704). It takes 16-bit / 44.1 kHz PCM over S/PDIF coaxial with no upsampling and no DSD. In Dmitriy Gubchenko’s voicing, Calliope is natural and organic, Persephone fast, Aphrodite dark and velvety, Artemis large-scale, Katerina strong in bass and attack. The stock output stage is solid-state; a tube stage on 6Zh4, EF14, EL8, AL4 or ML6, with a 5Ts4M rectifier, is optional.
• Hades (Class D amplifier) and Bellerophon (Class AB amplifier) carry a built-in DAC chosen at order time: the Cirrus Logic CS4397 as standard, or the TDA1540, TDA1541, TDA1547, PCM63, PCM1704, an R-2R ladder module or a programmable-logic module.
• Hyperion uses the CS4397 multibit delta-sigma chip for listeners who want USB input up to 24-bit / 192 kHz and DSD256.
• Prometheus is the programmable-logic converter: PCM is converted to DSD and passed through a window filter, with native DSD1024 and PCM up to 32-bit / 768 kHz.

Every converter is made by Dmitriy Gubchenko. Each unit ships with its own measurement report. If you are unsure which chip suits your system, write to us with your amplifier, loudspeakers and the music you play most.