Fixed Point vs. Floating Point

[crizdee]

Originally Posted by SimonN

Maybe somewhone who has stepped over to PTX can pull up a mix done all ITB on PTHD and post the two mixes here?

Hi,

No HDX cards are out yet! I guess someone could compare running a native system and all native plugs, but I think that would be quite difficult as an exact comparison.


Chris

[Greg Malcangi]

Originally Posted by lennieh

The antis say that if you need to use the extra range you get from 32-bit file formats then your gain structure is wrong. End of story. You can't directly present that file to your DA converter, so you have to dither it or apply gain reduction to it, so what's the point, you are just introducing more scope for error.

The pros would say it gives you added flexibility when mixing.

At the end of the day, do some blind listening tests and trust your ears

Originally Posted by MADIrouting

I can't say that I understand the 32-bit file process other than it is 24-bit with the 8-bits of headroom when saved.

It's not quite a simple as this.

1. 32bit float is not really the same as 24bit + 8 bit of headroom. That would be 32bit fixed. Basically 32bit float is roughly the same accuracy as 24bit with the additional bits used to define where the decimal point is (exponent).

2. While in theory 64bit float is roughly the same accuracy as 48bit fixed, the situation is a little more complex in the PT TDM world because not all of the signal path ITB is at 48bit, paths to and from inserts are 24bit for example. The 48bit summing mixer is not always 48bit either, there is a 56bit accumulator but depending on the number of channels you're running, the mix maybe invisibly sub-mixed (and reduced to 24bit) if spanning more than 1 TDM DSP chip.

3. The "antis" have a valid point in that floating point summing encourages incorrect gain staging but if you know what you're doing there are some nice workflow advantages to floating point if you're running lots of channels.

4. The "blind test and trust you're ears" mantra is usually good advice but should be taken in context. While you may hear sonic advantages of 32bit float over 48bit for one particular mix (or vice versa), that doesn't mean you have proved a rule that 32bit float is therefore always better than 48bit. There are also some potential disadvantages of floating point summing compared to fixed point summing, the aforementioned gain staging and the application of dither for example.

If I had to make a generalisation, I would say that I wouldn't expect 64bit float to sound worse than 48bit fixed and there are definite theoretical advantages to plugin routing and processing at 64bit compared to the current TDM architecture, depending on how it's implemented within PT and by the plugin manufacturers. I certainly wouldn't be expecting any night and day audible differences, very marginal differences which may only be noticeable in more extreme circumstances.

G

[Oblivion777]

Thanks for your replies! Can't say that I get everything 100%, but I'll re-read most of the posts and links several times (while getting back to Audio bit depth 101 1000x times).

In any case - I think - it is safe to say that in general and under normal circumstances the floating point will not sound a measurable amount worse than on the good-old HD (TDM) hardware. And that's good!

[Oblivion777]

By the way it would be great if an engineer from Avid would post some "arguments" (not that any defending would be needed) as to why they chose "the floating path" ... I know that the DSP chips operate in float environment, but I guess that when HDX was still on the drawing board there were several options/varieties to choose from and it would be great to know what arguments finally prevailed that lead to the decision.

P.S.: Admittedly I know that I am looking things way to narrow here as I am certain that the whole Fixed Point vs. Floating Point was not the prevailing argument here when the decision on the DSP chips was made.

[PaulV]

To reply to the original poster's question, I think the difficulty comes from confusion between 24-bit vs. floating-point storage and 24-bit vs floating-point processing. They are separate and independent from one another.

Storage: Pro Tools 10 can store recorded audio in either 24-bit or floating point files. Floating-point files provide an improvement only when storing a processed signal, such as with AudioSuite processing or mixing down tracks. They do not help with signals recorded from an audio interface. See this post for more details.

Processing: Mixing and plug-in processing are independent from the storage format you're using. Native and HDX systems use floating-point processing; HD uses fixed-point processing. You can produce great mixes with both of them. The main difference you will notice between the two is:

• Floating-point processing will not clip internally (such as when mixing between two busses), only at the output stage when going back to your interface (or when bouncing to a fixed-point file)
• With floating point processing, you don't have to worry about gain stages that bring a signal down really quiet. With fixed-point, you would lose signal quality if you bring the signal down far at one point, then back up later.

Also, the HDX and Native mixers use 64-bit accumulation for mixing, which will retain greater precision. Some people say they can hear the difference between this and 48-bit fixed point accumulation.

So the main difference with floating point processing is that your session's mix is much more forgiving about the gain staging, and won't clip or lose precision during routing between busses, sends, and submixes.

I hope that helps.

[jjnssn]

Originally Posted by PaulV

I hope that helps.

Yes. Thank you!

[Oblivion777]

Thank you Paul for your feedback. Really appreciate it and all the feedback from the users as well! All the posts have helped me to understand both the 32-bit float files and the 64-bit float mixing "stuff" ...right I realize that my thesaurus is a bit limited but hey English is not my native language...

[daeron80]

Some arguments/reasons have already been put forward by Avid and by some plug-in programmers. The videos Avid released about the time PT10 was announced contain some of them (search their channel on Youtube). This thread is astounding in its depth and the clarity of many of the official answers. If I had to summarize a bottom line off the top of my head, it would be in line with Greg's excellent summary above - computers eat float for breakfast, and doing summing in float allows the program to pass signals around inside a mix without converting them from float to fixed and back. Such format conversions can cause rounding and quantization errors to build up, and/or dither noise to cloud the details.

[carlone]

Originally Posted by corp

Wasn't there a time/argument when Avid/Digidesign conveyed that "fixed" was superior to "floating"....?

There was a time when computers were not fast enough to handle all that was needed for native processing/mixing so it wasn't even a fair argument. It's only been the last 6 maybe 7 years that floating point and host processing was a serious viable option for actual mixing ITB "compared to the fixed point TDM DSP option".

Just a thought

[daeron80]

There was also a time when common floating-point instruction sets didn't handle its mathematical idiosyncrasies well enough for audio DSP purposes. Fixed-point was the only way to insure reliable results. All that's been pretty well ironed out by now.