What is DSD, and what happens when you convert PCM to DSD
By Taiki NodaChecked with Sonir Bench 0.1.62 on 2026-10-15
DSD (Direct Stream Digital) stores audio as a 1-bit stream at a very high rate, made by a delta-sigma modulator that pushes quantization noise above the audible band. Converting PCM to DSD runs your music through such a modulator: it adds no information the source did not have, but it decides where the noise goes, how much the level has to drop, and which filter and modulator do the work instead of the ones inside your DAC.
The rest of this article explains each of those in turn, with numbers from our own modulator measurements, so that "DSD256 or DSD512" and "which modulator" become questions you can reason about rather than guess at.
What DSD is
PCM, the format of CDs and almost every streaming service, describes the waveform as a series of multi-bit numbers: 16 or 24 bits per sample, 44,100 to 192,000 samples per second. Each sample says "the signal is here."
DSD uses one bit per sample. Each bit only says "up" or "down." A single bit cannot describe a waveform on its own, so DSD runs at a rate far above anything PCM uses, and the density of ones versus zeros over a short stretch of time carries the signal. Filter that stream with a low-pass filter and the music comes back out.
The rates are named as multiples of 44.1 kHz:
| Name | Rate | Multiple of 44.1 kHz |
|---|---|---|
| DSD64 | 2.8224 MHz | 64x |
| DSD128 | 5.6448 MHz | 128x |
| DSD256 | 11.2896 MHz | 256x |
| DSD512 | 22.5792 MHz | 512x |
DSD64 is the rate used on SACD. The higher rates exist mostly as download formats and as targets for PCM-to-DSD conversion on a computer.
Delta-sigma modulation in one paragraph
Turning a fine-grained signal into one bit throws away a huge amount of resolution. That error (quantization noise) has to go somewhere. A delta-sigma modulator feeds the error back and subtracts it from the next samples, so that the noise is pushed out of the low frequencies and piled up at high frequencies. This is called noise shaping. The audible band stays clean; the space between roughly 20 kHz and half the DSD rate fills up with noise. The higher the DSD rate, the more room there is above the audio band to put that noise, which is the main reason higher rates exist.
The shape of that push is set by the modulator's noise transfer function, and the steepness of it is described by the modulator's order. A higher order digs the audio band deeper.
What happens when you convert PCM to DSD
"Upsampling to DSD" is two steps:
- Interpolation. The PCM is raised to the DSD rate with an oversampling filter. This is the same kind of filter any upsampler uses, with the same choices (phase, length, where the stopband begins).
- Modulation. The high-rate signal goes through a delta-sigma modulator and comes out as a 1-bit stream.
What you get
- The filter and modulator are chosen by you, not by the DAC. Most modern DACs are themselves delta-sigma designs: they oversample PCM and modulate it internally with whatever the chip designer chose. Sending DSD hands the DAC a stream that has already been through those stages on the computer, where far more processing power is available. How much of the DAC's own path is skipped depends on the DAC's design, and varies from model to model.
- A consistent output format. Whatever the source rate, the DAC receives the same DSD rate.
What you do not get
- No new information. A 16-bit / 44.1 kHz file converted to DSD512 still contains only what the 16-bit / 44.1 kHz file contained. The conversion can be very clean, but it cannot add detail.
- No free level. A modulator cannot be driven to full scale. Every delta-sigma modulator has a maximum stable amplitude (MSA): push the input past it and the feedback loop stops tracking the signal and runs away. So the signal has to be turned down before it enters. In our modulator, the reduction needed at the four orders we offer ranges from 3.3 dB (3rd order) to 6.4 dB (6th order). That is not a volume control and not a loss of resolution in itself, but you will hear the output as quieter, and if you compare DSD against PCM without matching levels, the louder one tends to win.
- No escape from the noise. The shaped noise above the audio band is part of every DSD stream. You can choose where it sits; you cannot remove it.
Order and stability: what you are choosing when you choose a modulator
Players that convert to DSD often let you choose the modulator. The names differ between products, but every modulator choice is a trade between the same few properties. We measured them on our own modulator design, at DSD256 with the out-of-band gain fixed at 1.5:
| Order | Effective SNR | Level drop needed |
|---|---|---|
| 3 | 132.6 dB | 3.3 dB |
| 4 | 159.9 dB | 4.8 dB |
| 5 | 186.2 dB | 5.6 dB |
| 6 | 209.8 dB | 6.0 dB |
These were measured over a window of about 82,000 samples. The level drop actually applied comes from a stricter test over a million samples and is slightly larger, up to 0.5 dB more at 6th order (6.4 dB).
Two things move together: raising the order raises the SNR, and in exchange the level has to drop further.
The SNR column is not what decides
209.8 dB looks dramatically better than 132.6 dB. It is not, in practice. The analog side of a good DAC has a noise floor around 120 dB, so every order from 3rd upward already sits above anything the analog output can show. Past that floor, the extra SNR exists only inside the computer.
There is a lower limit, though. At 2nd order, the effective SNR in our measurement is 88.1 dB, which is below the 96 dB of a CD. That is why we do not offer it. Sonir Bench defaults to 3rd order, which needs the smallest level drop, and lets you choose up to 6th.
Higher order is harder to keep stable
The folklore says that modulators above 5th order are hard to make stable. In our measurements the real condition is the level. Fed at nearly full scale, only low orders survive. Turned down by a few dB, designs up to 6th order run without trouble. The MSA falls as the order rises, and that falling MSA is exactly the "level drop" column above.
One practical warning from testing: a modulator can look stable for a while and then diverge. With a short test window of around 50,000 samples, one setting we tried passed and then blew up at sample 264,000. We test stability over a million samples for this reason.
Cost
People often assume the modulator is what makes DSD512 heavy. In our tests on an Apple M1, the whole conversion stage (interpolation, level drop, modulation, packing the bits) for stereo DSD512 took about 38 to 48% of one core, and stereo DSD256 about 19 to 24%. The order barely moves the cost: at DSD256, 3rd order took 19.0% and 6th order 23.7%. The limit is the feedback chain itself, which has to run one sample at a time, not the arithmetic.
DSD256 vs DSD512
Doubling the rate doubles the room above the audio band. For a modulator of the same design, that means a deeper audio band, or the same depth with more room to spread the noise. In our design the level drop moves very little with the rate (by 0.1 dB at 3rd order).
What changes more is what has to be true of your equipment:
| DSD256 | DSD512 | |
|---|---|---|
| DSD rate | 11.2896 MHz | 22.5792 MHz |
| PCM rate needed for DoP | 705.6 kHz | 1411.2 kHz |
| CPU in our tests (stereo, M1) | 19–24% of one core | 38–48% of one core |
If your DAC accepts DSD512 over DoP, there is no technical reason not to try it. If it does not, DSD256 is not a compromise you need to worry about: the audible band is already far below the DAC's analog floor at either rate.
Building DSD from 48 kHz-family material
DSD rates are multiples of 44,100 Hz. Music at 48, 96 or 192 kHz does not divide into them evenly. There are two ways to deal with that:
- Convert to the 44.1 kHz family first. Where the conversion happens matters a lot. Done at the source rate, a 48 → 44.1 kHz conversion needs a steep filter right at 22.05 kHz, uncomfortably close to the music. Done after upsampling, it has room to spare: our converter goes from 768 kHz to 705.6 kHz (a ratio of 147/160), where the music is already band-limited to 24 kHz and the filter has hundreds of kilohertz to work with. In our measurements, the response stays within 0.01 dB all the way to 20 kHz, and the rejection is around 118 to 120 dB.
- Send 48 kHz-family DSD as it is. DSD64 in the 48 kHz family runs at 3.072 MHz instead of 2.8224 MHz. Some DACs accept it, but you cannot tell from the PCM rates a DAC lists whether it will interpret DoP at those rates. Only the manual or a test will tell you.
If a DAC receives DoP it cannot lock onto, it treats the stream as ordinary PCM. When we tested this, the result was not full-scale noise: it measured -26.6 dBFS, almost all of it a tone at the top of the output band, far above hearing. It is not loud, but check that your DAC supports DoP before using it.
What a DAC needs for DSD512 on a Mac
macOS has no native way to send DSD to a DAC, so on a Mac DSD always travels as DoP (DSD over PCM). DoP packs 16 DSD bits into the lower 16 bits of each 24-bit PCM word and marks the top 8 bits with alternating values (0x05 and 0xFA) so the DAC can recognize it. The PCM rate is therefore always one sixteenth of the DSD rate. DoP itself is a separate topic; here is what it means for DSD512:
- The DAC must accept 1411.2 kHz PCM. That is twice the 705.6 kHz that many high-rate DACs list as their top. Open Audio MIDI Setup, select the DAC and look at its formats. If 1411.2 kHz is not there, DSD512 over DoP is not possible on that DAC.
- The DAC must recognize DoP at that rate. Spec sheets often list a native DSD maximum and a DoP maximum separately. On a Mac, the DoP figure is the one that applies.
- macOS must open the stream. Core Audio opens streams at 1411.2 kHz: Sonir Bench plays DSD512 over DoP from a Mac mini to an external DAC.
As an example of the first point: the Topping E30 II Lite we test with lists rates up to 768 kHz. 705.6 kHz is on the list, so a Mac can send it DSD256 over DoP. 1411.2 kHz is not, so DSD512 is not available, and the stream refuses to open rather than playing something wrong.
FAQ
Is DSD better than PCM?
Neither format is better in itself. A DSD stream converted from PCM carries the same information as the PCM it came from. What converting changes is which filter and modulator do the work, and whether that sounds different on your system depends on how your DAC handles each input.
Should I use DSD256 or DSD512?
Use the highest rate your DAC accepts over DoP and your computer runs without dropouts. Both put the audio-band noise far below what a DAC's analog output can reproduce. On a Mac, DSD512 needs the DAC to accept 1411.2 kHz PCM.
What does the modulator order change?
Higher order digs the audio band deeper and needs a bigger level drop. From 3rd order up, the audio-band SNR is already above a DAC's analog floor, so in practice the difference you pay for is the level drop.
Why is DSD output quieter than PCM?
A delta-sigma modulator becomes unstable near full scale, so the signal has to be turned down before modulation. In our design that is 3.3 to 6.4 dB depending on the order. Match levels before you compare.
Can I convert 48 kHz music to DSD?
Yes. Convert it to the 44.1 kHz family at a high internal rate, where the conversion does not touch the audible band, or send it as 48 kHz-family DSD if your DAC is documented to accept it.
What Sonir Bench does here
Sonir Bench (0.1.62) converts what your music app plays into DSD64, DSD128, DSD256 or DSD512 and sends it to your DAC over DoP. In the Output Settings card on the Engine screen, set Output format to DSD (DoP) and choose a DSD rate; the exit rate follows from it. Under Order you choose 3rd to 6th order (3rd by default), with the measured Effective SNR and Level drop shown for each next to the DAC's analog floor, and the level drop actually applied is always on screen. Convert to 44.1 is on by default for 48 kHz-family material. For the steps, see Choose how Bench feeds your DAC: higher rates or DSD.