MOTUnation.com

You are falling into the same boat as I was about resolution, you are not gaining resolution inside of DP with 32-bit, you are gaining range. At first I didn't understand it until it was explained quite fully. All 32-bits are not created equal.

If you record something with a 24-bit interface, it doesn't matter where you place that 24-bit record into a 32-bit dynamic range, because the noise scales with you up or down. So starting at 24-bit and going to 32-bit if we were talking about resolution wouldn't push your noise floor down, that would have to happen on the actual recording device. If you can do that, then yes, you'd be able to theoretically push the floor down further IF the noise floor was lower on that device. That completely depends on the converters. Once you are in the digital domain, noise floor isn't something that is going to increase, you're digital at this point.

But, like I said, In DP the 32-bit is actually including an 7-bit (I believe) exponent that places your 24-bit data into the 32-bit window. So any noise will be the same even if you went to 64-bit. I tested this extensively as I was wondering why the internal bounce didn't let you do 32-bit.

At 32-bits, your dynamic range is 1500db. This is MORE than adequate for mixing in DP. Again, because you are not gaining resolution in your signals. There is no round-off error except at the very end when you go back from 32-bit to 24 (or 16). There are some very detailed threads about this on here and on OSXAUDIO, and I'd suggest reading them. They certainly opened up my eyes.

Last, comparing 16-bit to 24-bit as opposed to 32-bit to 64-bit doesn't make any sense. 16-bit is going to be below the dynamc range of your equipment, so of course you are going to hear a difference. I'd bet money that given almost any project most people have worked on, taking DP from 32-bit dynamic range to 64-bit will buy you nothing, because you aren't using 1500 dB of range as it is. 32-bit is actually quite huge.

Again, wrap your head around the way the 32-bits are used in DP. Your 24-bit input is adjusted by the exponent in the range. This is how it was explained to me, and all my tests I did with bouncing to 24-bit and comparing to original values using inverse sums proved it to be true, and I was VERY thorough in this.

well, I‘m not a 100% sure, but there were some basic comparisons of signal summing in DAWs some time ago. And the results even when just summing up were audibly different. so they said. So it‘s not just about dynamics.

As far as I can see, the 24 bit-signals stay 24 bit and are not being converted but summed up and processed with higher precision. May be that you can‘t tell the difference between 32 bit and 64 bit, I honestly have not tried to do a comparison and yes, part of the 64 bit-hype is just to have us buy stronger computers.
But I don‘t think it‘s all about nothing (which is to say I don‘t know).

The difference between PCM and analog is that still we draw in steps. The higher the resolution, the smaller the difference to the original signal. It‘s a shame, DSD didn‘t make it into the recording domain, I was really pleased with the recordings I bought. The difference is that DSD doesn‘t measure absolute values, but differences which leads to rectangles where PCM does steps- a step closer to the source signal.

I don‘t think that high bit mixing engines just add empty bits on one end which would lead to higher resolution in noise or add headroom where it‘s not needed because the signals won‘t reach into that area which just a minute ago meant clipping...
Maybe I‘m wrong and it is in deed all about head room?

Summing two phase coherent sinus signals leads to twice the amplitude meaning + 6dB Full scale. So the summed Signal would extend the dynamic range by 6dB, dropping the noise floor by the same amount? Or giving at least 6 dB space before noticable quantization noise? And what about summing three or four signals?

Maybe the 24bits are somehow "stretched", like in an expander?

Where are the pros when one needs them?

oh and a word on the 16 bit- 24 bit-thingy:

16 bits give you a theoretical dynamic range of 93 dB, of which about 74 are really usable. Most productions are even further limited to somewhat around 50 dB. Keep in mind that you‘d have to listen at a peak of at least 60 dB SPL (in case your listening environment is nearly quite; 10dB isn‘t much at all) to be able to hear the loudest and quitest detail of a recording with a dynamic range of 50 dB.
Well I for one would state that this wouldn‘t be at all a reason why the different resolutions should sound differently if it wasn‘t for the higher precision in summing the signals as well...

And furthermore, I couldn‘t listen at such high levels, due to my flat being built of paper or whatever may be the reason for hearing so much of the neighbours I really never wanted to know

Hehe, I hear you on the listening levels. I wish I had the link handy but I went through a lot of the same thoughts you are going through now, and a LOT of people went to great detail to set me straight.

I had originally thought the 24-bit signal was stretched into 32-bit, but I was wrong, and this was from information from Motu specifically. This is what turned into my revelation on what exactly they were doing with DP and 32-bit.

For instance, if you are in the red inside tracks that are internal (say on busses) you aren't clipping. It's that last track going out that needs to stay within the green. (or to a bounce). I tested this which helped me understand what those 32-bits were all about, and that is range. For instance, if you take 2 signals and add them together, you need but 1 bit extra to show the sum. If you take 2 signals you'll need 2 bits extra, but also, even if you added three signals (assuming a maximum value). This is why 1500dB of headroom on 32-bits shows that you are above any beyond anything you are really going to need.

And there is the difference, headroom.

Now the question of resolution is different. DP doesn't give you 32-bits of resolution. It works at 24-bit or 16-bit, depending on your hardware. What has been explained to me is that if you are starting at 24-bit, since you are being placed into the 32-bit field, your noise floor in your 24-bits remains the same. This would not go up or down if you were at 64-bits of headroom because that extra headroom isn't being used. BUT at the same regard, if you stretched your data to fit in 32-bit if it were a resolution difference, you're noise would be amplified as well. As I am sure you are aware.

This is why it makes complete sense to me that DP had to go to 32-bit internal, wouldn't make and sense not to.

Most of the hype you read about 64-bit processors has to do with memory space only. This is important if you have lots of RAM and are loading big sample sets or recording to memory. For some people, especially orchestral, these higher memory extentions are going to be a godsend, eliminating costly streaming from disk.

For people like me that do a lot of real time processing, 64-bit isn't buying me anything. I'd rather run 32-bit code if it's going to be faster, and it is on certain architecture, like the opterons. But, as has been mentioned, it depends on a lot of what is going on inside the processor. It depends on the cache sizes. And to quote an excellent example of poor design, the Itanium from Intel executed 32-bit code very poorly, and we see where that got them.

Floating point math inside the processor has been 64-bit and higher for a long time. Some plugins might use this extra resolution internally, but without specifically getting the information from them, you won't know if they are expanding the resolution to eliminate round-off errors, or expanding the range to eliminate clipping issues.

Which of course, is the same confusion I was hit with.

I am not sure which Ozone is using inside, and am very curious.

But as for round-off errors with DP, you don't get them. The test I did was feed a 24-bit signal through 5 reverbs and record the output. Then, at each stage, I bounced the track to disk at 24-bit and fed it to the next track. Hoping to see quantization happening at each stage. At the end I inverted one of the signals and added it back to the one I recorded in one pass. There was zero output. There was no round-off error in either except at the very end which was the same. This all made sense to me then.