32 bit DAC??? Is that even possible?

[Berni]

Most audio DACs have no extra control bus to even have registers. Usually its just a few extra pins that can be set high or low to enable or disable some features. On the lower pin count ones you don't even get that. Its just I2S in and analog out the other side.

At the same time you can buy a 24bit stereo audio DAC from name brands like AKM, Texas, Cirrus for $1.20 to $2.00 on digikey. So there is not really much room for saving money by going with less than 24bit. They have good specs and will not be the limiting factor once paired up with cheep opamps in the analog stages. Because lets face it if you bought a $1.20 DAC chip, then you are not going to pair it up with a pair of $2 opamps per channel. Whatever is producing the audio signal likely also has a hardware multiplier so volume control is literally a few CPU cycles to do. On the other hand a external digital pot for volume control is going to cost at least as much as that DAC while doing a poor job designing it might result in I2C bleeps and bloops leaking into the audio output.

Its the same argument as 32bit ARM MCUs doing stupid simple tasks that your oldschool classical 8 banger could also easily do. Yes you don't need a 32bit ARM core to do the job, but at this point the cheep tiny ARMs now cost similar as name brand 8bit MCUs cost. So you might as well just go for 32bit and enjoy the luxuries such as a CPU core that was actually designed for running C code rather than hand coded assembler, actual interrupt vectors, proper debug functionality... etc.

[magic]

There surely are audio DACs with volume control over SPI or similar interface, though the examples I could name aren't cheap

And I think they used digital filter for it, not PGA.

[Berni]

Actually, even only doing 8 bit data manipulation, 32-bit CPUs can yield tighter and faster code as well. With longer instruction words, you can encode in more conditions and immediate numbers, which will cost more cycles on a 8-bitter, thus you can have higher throughput per instruction cycle.

Yeah the 32bit also provide a lot of efficiency in computation since we very often work with numbers that don't fit into 8 bit nicely.

But my point is more for applications where performance doesn't matter at all. Like some thermostat with a 7 segment display and a few buttons, a electronic combination lock, an remote control receiver for a garage door... etc. Things that almost any crappy slow MCU could do even when running at 1/10 its rated clock speed. Yet these days it makes sense to use cheep little 32bit MCUs for these dumb tasks since they cost about as much as any other name brand MCU.

Yes i know about dirt cheep Chinese MCUs that literally cost a couple of cents each and tend to be a 8bit MCU (pretty basic even at that). But unless you are doing something that is really high volume and cost sensitive then people from the west are going to stick to western name brand MCUs since they know how to use them, where to buy them and can read the documentation without a translator. These western 8bit MCUs are nowhere near as cheep.

[Ed.Kloonk]

If you ever wanna see some cool audio software engineering, check out the MAD mp3 decoder library. 24-bit output, integer-only code. No floats.

Nearly 20 years old now.

[Jan Audio]

Pretty much the only common use for floating point audio is in audio editing and music creation software.

Things change.
You now have 64 bit floating point in STM32 ( sold out )
It should be faster doing floating point math since you have no shift.
*i dont have STM32 due to package restrictions, i can see the advantage of 64 bit floating point in 1 instruction.

[Jan Audio]

Suppose they make a 32 bit DAC on -15 to +15 volt, the s/n ratio is improved right ?
Or maybe high voltage for the best quality.

[Ed.Kloonk]

Suppose they make a 32 bit DAC on -15 to +15 volt, the s/n ratio is improved right ?
Or maybe high voltage for the best quality.

There was a use case theory ages ago, among us HQ audio snobs, where that concept could be used in active speaker modules. Sounded good, in theory. What we got instead were active speaker modules with sloppy old 16-bit DACs, class-d amps and really terrible power supplies.

[bson]

while attenuation in firmware costs 2kB of code space.

With a hardware multiplier and no sigma delta dithering happening, maybe a few tens of bytes of code .

True, but you also need some way of controlling the volume.

[Bassman59]

while attenuation in firmware costs 2kB of code space.

With a hardware multiplier and no sigma delta dithering happening, maybe a few tens of bytes of code .

True, but you also need some way of controlling the volume.

Meh, that's what knobs are for!

[Humanoid]

I work with audio a lot, so here's my 2¢ (or maybe 10).

I use a pro audio interface (MOTU 1248) and don't need a dedicated ADC or DAC for my work. Most modern ADCs and DACs on interfaces handle 32-bit conversion, if they cost more than $20. 32-bit 192k is pretty standard specs even for mid range desktop interfaces now with 2 ins and outs. I personally see no benefit to using those higher sample rates as they can add in harmonics and other garbage above 20k during a mix which brings the entire level of the mix up. Some engineers swear the highs are clearer at 192 and 384, but I have no need for it. Tokyo Dawn Labs has an old plugin specifically made for filtering those ultrasonic frequencies out of a mix. 32-bit could be useful though if you need it.

In the late 90s when I went from 16-bit to 24-bit, there was a big difference because the signal to noise ratio is so much better with 24. No longer did we need to record hot signals. These days I try and record levels around -18dB to -12dB so I have plenty of headroom when mixing. 24-bit to 32-bit the noise floor is usually the same. It's the massive headroom over digital 0 that differs. If you hear noise in your 24-bit recording at rational listening levels, it's probably interference, or you're recording too low and raising the gain too much, or your monitoring signal chain is noisy, or maybe you have grounding problems (typically hum though).

I used to use 96k 24-bit back when many plug-ins didn't upsample because I could hear a difference in the output, but these days pretty much everything upsamples to some ridiculous degree to avoid aliasing and other artifacts and then downsamples on the fly, unless you are using some old plug-ins that aren't being updated. I can hear no difference in output anymore so I stick to 48k sample rate as a standard unless a client demands a higher one. It keeps files smaller, avoids the garbage above 20k, is the standard for video and cinema, and drastically frees up your CPU to use way more plugins. I can do mixes on my modern 10-core CPU using dozens and dozens of tracks with tons of plugins that would need accelerator hardware in the past. My heaviest mixes usually don't pass 50% CPU usage. God bless multi-core.

As for bit rate (or depth if you prefer), 32-bit float I find to be beneficial when recording because of the extra headroom. You can record values past digital 0 to avoid clipping and then bring down the gain when mixing (or surgically with bite volume or in a 3rd party app like RX or Audacity). I think the headroom surpasses the decibel levels of sounds we can even make on earth. I find it useful when recording voices and sound effects, because you can get spikes if someone goes closer to the mic and ruin a take. If you like to record hot, 32-bit is also your friend for the same reason, you can always mix down to 24 once things are tamed and most DAWs these days let you mix different bit rates in the same sequence if you really want to.

In pro studios, 24-bit is the standard because they won't record hot and most engineers won't mix to the ceiling. Some swear by 88k because it's double 44k used for CDs, and supposedly there will be no rounding errors when downsampling. You may find 32-bit being used in the final mixing phase at some studios so if something goes over the mastering engineer can tame it and not be dealing with a clipped square wave. It all depends on the workflow. These days people send stems to mastering engineers sometimes so nothing will be clipping (in theory). If you have the hard drive space and want to avoid having clipping problems, use 32-bit float. If not, 24 is fine. There's even free plugins you can drop on your tracks and buses to check if there are inter-sample peaks the meters don't register.

I personally do all my recording and audio mixes at 48k 32-bit float for A/V projects to avoid clipping and because Adobe After Effects does not natively export 24-bit. There is no audible difference between 24 and 32, it's merely so I can avoid clipping and don't have to change formats when doing intermediate renders and when working with different clips, or different audio sources and music. I keep everything at 32-bit 48k throughout the entire production phase so regardless of what software I have to use, no problems arise. If I have to convert sample rate or bit depth, I do it at the start using iZotope RX7 because it has one of the best software converters. Some people prefer Weiss Saracon or Saracon and RX (or a few others), though RX is price tiered and has more tools.

If you work on A/V for the web, YouTube even accepts 48k 32-bit float linear PCM embedded in video (even if the video is compressed as H264), so you can let them do the final audio compression after upload for the best audio quality and avoid double compression.

So basically, if you have a good "pro-sumer" or professional level interface, the bit depth/rate you choose will really depend on what you're doing and most likely it will matter most in the software when recording or doing a final mix. All the well-known manufacturers use great preamps and their ADC and DAC is top notch. Even modern soundcards on motherboards handle 32-bit audio playback these days. Use 32-bit for practical reasons, not because of a sales pitch.

[JohnnyMalaria]

while attenuation in firmware costs 2kB of code space.

With a hardware multiplier and no sigma delta dithering happening, maybe a few tens of bytes of code .

True, but you also need some way of controlling the volume.

Meh, that's what knobs are for!

As long as they go to 11 and you can make everything louder than everything else.