32 bit Float question for Magic Dave

[James Steele]

guitardood and Gravity Jim: I think BOTH parties need to sort of check the egos a little bit. I'm deleting a couple of the messages. I don't have time to jump in and sort out who started it, etc.

I will say one thing. I often jump into a thread and make "not happening for me" posts just to give some balance to a topic and make sure that outsiders reading the forum don't get the impression that DP is just some totally messed up DAW that doesn't work for anyone. It works for many of us, but forums like this naturally attract complaints. That's why I set up the whole "Criticism" forum (which perhaps this topic belongs in) so that those of us for whom DP is working fine had a refuge, where we could discuss and share ways to actually use the DP, instead of being barraged by thread after thread about what isn't working right for some people.

[magicd]

Geez my ears are burning!

There is no good reason to record at 32 bit float. The drivers for the interface are sending in 24 bit integer samples so initially when recording only the top 24 bits of the sample will be used.

DP uses 32 bit floating point processing internally (some plugs can go up to 64 bit float). It is possible to internally exceed the maximum gain that could be described by a 24 bit sample, so you have to be careful when adding gain inside the mix engine, in terms of what the final output will be.

Playback through the audio interface has to be brought down to fit into a 24 bit sample because again that's what the D/A converters are looking for. Recording a 32 bit has no affect on A/D or D/A conversion. Internal mixing does use 32 bit float and can exceed the 24 bit range. That's why you use a master fader and make sure you are not clipping at the output of the master fader.

So what good are 32 bit float files? IMO they are very specialized. You can bounce or freeze to a 32 bit floating point file. If you do so you can extend the dynamic range of the signal. My understanding is that a 24 bit sample can describe 144 dB of signal. A 32 bit float signal has around 1500 dB of available dynamic range.

So you could bounce or freeze to a 32 bit file but you still can't play that file back without attenuating playback volume to within the 24 bit integer range (because of the D/A converters).

For reference, humans are generally understood to have no more than 120 dB of dynamic range in their hearing and I personally probably have about 70 db of range left (don't tell my mastering clients).

So the only practical use for a 32 bit float file is to store audio that has exceeded the 144 dB range of the 24 bit sample.


Shooshies first post described distortion. That may be because of gain staging or exceeding the 24 bit output to D/A. I do know that all stock MOTU plug-ins can handle a 32 bit float signal because they are designed to do so. I don't know if that's the case with any particular third party plug-in. Also keep in mind that some plug-ins are gain dependent in terms of what they do to the signal. If there is any question that distortion is coming from a plug-in, take the plug-in out of the signal path and see what that does.

Last, keep also in mind that 32 bit float files are twice the size of a 16 bit file and a third larger than a 24 bit file. You are asking the computer and drive to do that much more work. Just like we can kill a system be eventually overloading it with tracks and plug-ins, you'd get there faster with 32 bit float files. Like everything else we can test by simplifying the system and seeing at what point the signal cleans up.

Hope that helps!
Dave

[guitardood]

Last, keep also in mind that 32 bit float files are twice the size of a 16 bit file and a third larger than a 24 bit file. You are asking the computer and drive to do that much more work. Just like we can kill a system be eventually overloading it with tracks and plug-ins, you'd get there faster with 32 bit float files. Like everything else we can test by simplifying the system and seeing at what point the signal cleans up.

Thanks for the great info!!

Question:With the latest pregen engine, would this also apply to the pregen'd audio files, if any, for Aux busses? Would explain a lot in my particular circumstance.

[stubbsonic]

Just out of curiosity (and some ignorance thrown in for good measure)-- I'd like to understand the difference between RESOLUTION and GAIN where the bit format is concerned.

When DP is using a 32-float, does the factor of it being "float" mean that it is interpreting the 32 bits to have a higher resolution as opposed to having 1500 dB of gain?

To ask a different way, if I have 16 bit audio normalized and the same audio as 24 bit normalized, I expect it to look same on the meters, but the 24 bit file will just have more resolution. I don't know what has to happen behind the scenes with the DACs to make that so. When we have 32-float files, should we expect them to look different on the meters (and clip internally?)

[oshtakuta]

Question for all After finishing the project (MIDI virtual instruments and recorded in a studio live tracks), which format should I choose for bouncing the STEMS: 24 bits or 32-bits? My goal is to send he project to a sound engineer for mixing. Also, should I add dithering? Please give a definitive simple answer for DP only (e.g. Logic will bounce everything in 24-bit format and this is a bit confusing).

[HCMarkus]

Question for all After finishing the project (MIDI virtual instruments and recorded in a studio live tracks), which format should I choose for bouncing the STEMS: 24 bits or 32-bits? My goal is to send he project to a sound engineer for mixing. Also, should I add dithering? Please give a definitive simple answer for DP only (e.g. Logic will bounce everything in 24-bit format and this is a bit confusing).

I'd say 24 bit, no dither.

[oshtakuta]

Audio from DP 32bitFloat !!!!!!!!!!! O My God

[terrybritton]

Just out of curiosity (and some ignorance thrown in for good measure)-- I'd like to understand the difference between RESOLUTION and GAIN where the bit format is concerned.

When DP is using a 32-float, does the factor of it being "float" mean that it is interpreting the 32 bits to have a higher resolution as opposed to having 1500 dB of gain?

To ask a different way, if I have 16 bit audio normalized and the same audio as 24 bit normalized, I expect it to look same on the meters, but the 24 bit file will just have more resolution. I don't know what has to happen behind the scenes with the DACs to make that so. When we have 32-float files, should we expect them to look different on the meters (and clip internally?)

My understanding is that in any format, 0DB is still all ones or "full-on", and the resolution goes in the "quieter than full-on" direction, which is why 24-bit is better for reverb tails and piano string fades, and we have all that room to use -12db as our peak value. (The quieter the content is, the fewer bits it uses.)

So, yes, there is more resolution, especially for the quiet sections which, in 16-bit, become very low bit "words", while staying "longer words" with 24-bit even at the same low "volume" (all things being relative to "full-on"). 32-bit is 24-bit with room left for even more fine-tuning between levels with the mantissa/decimal fractions added. So, that being the case, higher resolution is what you are getting there, too.

24-bit is entirely adequate really. 16-bit has trouble with the quiet stuff capturing the subtleties. But if you have low dynamic range in the recording and no super quiet events, then 16-bit is all you need. (Plus, for the final mix, you usually are going to compress the dynamic range anyway for the 16-bit final to keep those tails out of the listening environment's noise floor at average listening levels!)

0db is still "full-on", though! Don't go over that!

So, to sum up:
That increased "dynamic range" value does not mean it makes stuff louder, practically speaking. It means there are more numerical segments between each volume level, especially at lower volumes (which suffer noticeably at 16-bit - the lower the volume, the fewer bits are used to represent that volume). What 16-bit audio gave you 4-bits to reproduce, 24-bit gives you more relative to full volume, so the resolution is higher between those subtle volume differences way down there.

4 bits = 16 values
5 bits = 32 values
6 bits = 64 values
7 bits = 128 values
8 bits = 256 values

Terry

[stubbsonic]

That was my albeit vague understanding.

I was inferring from Magic Dave's post that something about 24 bits or 32-float made it easier to exceed gain limits than say 16 bit. And that does seem to be what Shooshie was describing.

[terrybritton]

That was my albeit vague understanding.

I was inferring from Magic Dave's post that something about 24 bits or 32-float made it easier to exceed gain limits than say 16 bit. And that does seem to be what Shooshie was describing.

Right - because "full-on" can represent such a high value in 32-bit, and even "half-on" is very high in that world, the plugin may be designed only to work well with much lower values. However, when you crank things down so much to match to the plugin's expectations, you have returned to the same resolution as you had with 24-bit anyway, so there is no advantage... quiet sounds end up with the same number of bits. Keeping in mind that all-ones is "as loud as one can get" and everything with fewer ones is measured by a negative-dbs, the relativity factor is "-10 db relative to what?"

That has to be the most confusing part of this!

Long and short of it - 32-bit float is handy for archiving to hand off for a mix and for internal processing, 24-bit is great for recording as it is within the range that most plugins can understand, and 16-bit with slightly compressed audio is ideal for listening environments with higher environmental noise floors than the 16-bit digital "noise" floor is (-96db on the meters).

Like, most of the planet!

Terry

[stubbsonic]

Oooph. This is confusing. If we thought of full-on as "1", or "1.0000.....000" ;
and half of full as either .5 or .500000 or .500000000000000000000000000000000,
the values would be interpreted as equal, in terms of level-- with more available points in between in the case of 32-float.

The level interpretation would be basically the same. That's why this topic is so confusing. Either the extra bits represent added "digital gain" or they represent added "digital precision" or some combination of both.

[terrybritton]

Oooph. This is confusing. If we thought of full-on as "1", or "1.0000.....000" ;
and half of full as either .5 or .500000 or .500000000000000000000000000000000,
the values would be interpreted as equal, in terms of level-- with more available points in between in the case of 32-float.

The level interpretation would be basically the same. That's why this topic is so confusing. Either the extra bits represent added "digital gain" or they represent added "digital precision" or some combination of both.

Right you are, and that struck me a moment ago that I had ignored the fact that 32-bit float IS 24-bit but with a mantissa (that is, it is not 32-bit non-float, which would have that super crazy high dynamic range) - MD threw me for a second. But.... Hrmmm

Terry

[oshtakuta]

https://www.gearslutz.com/board/high-en ... yours.html

[oshtakuta]

"Holy cow, thanks for this thread! Like probably most people, I had no idea you were supposed to add dither to every DAC output. Makes sense. Got some TPDF on there and it definitely sounds better!" https://www.gearslutz.com/board/high-en ... rs-17.html

[ccrane]

Wow, what an intense thread.

I've been using 32 bit float to record all of my VI's for a while now. The rational behind that was to preserve every bit of digital resolution that the VI was likely putting out.

I've had times in the past where after freezing every instrument, the mix sounded 'different'. Obviously, in the 16 bit days, that made sense, but even at 24 bits, I was unconvinced that I had not lost some depth when all tracks (100+) were frozen at 24 bits and played back.

Additionally, it seems that recording stems and final mixes would benefit from 32 bit float.

But, it appears that many of you, including Magic Dave are saying, or inferring, that there may be no benefit of freezing, even a VI, (that is possibly operating in 32bit float) to a 32 bit float file.

Possibly related to Shooshie's original post, I also have been getting unexpected peaks on the Master fader that will not go away, even with brick walling. But in those instances, a restart always clears the peaks, and those peaks only happen when I am switching between several large projects in a day (although with DP restarts in between).