Login / Register
 
Does 64-bit summing yield noticeably better output?!
New Reply
Subscribe
geartommy
Thread Starter
#1
29th February 2012
Old 29th February 2012
  #1
Gear interested
 
Joined: Apr 2011
Posts: 22

Thread Starter
geartommy is offline
Does 64-bit summing yield noticeably better output?!

With my limited mathematical knowledge I understand 64-bit summing should yield better results before dithering/truncation, but is it noticeable?!

With lots of plugs and automation on dozens of tracks, I guess the end number representing my mix is quite nasty looking and... loooong, once it's calculated. Obviously, it makes sense to use the highest precision engine (i.e. working with the longest numbers), but is that going to help if I'm using 32-bit plugs for example?!

Been using Logic, which is supposedly having a "mere" 32-bit summing engine, but apparently Avid are touting 64-bit for summing and 32-bit for processing. Other companies like Cakewalk and Cockos have been marketing their 64-bit summing engines for years.

Not wanting to believe the hype, what's deal here? Is 64-bit summing yielding better masters?

P.S. Don't have Pro Tools 10, so can't do any tests, unfortunately.
#2
29th February 2012
Old 29th February 2012
  #2
Gear addict
 
Joined: Jun 2011
Posts: 304

Gear Geek is offline
It is most relevant in plug-in processing and only has and advantage with plug-ins that are using 64 bit processing.
#3
29th February 2012
Old 29th February 2012
  #3
Lives for gear
 
Joined: Aug 2010
Posts: 3,283

-tc- is offline
Quote:
Originally Posted by geartommy View Post
P.S. Don't have Pro Tools 10, so can't do any tests, unfortunately.
Nonsense

Luckily for us all, the chaps at Cockos have catered for this very scenario. Download the trial of reaper if you don't already own it.

Do a mix and render it. Then open project settings, click on the "Advanced" tab and change the "Track mixing bit depth" to 32 bit float or whatever takes your fancy, and render again.

Then abx. Then put it back to 64 bit float (cz ... why not), and forget all about it
#4
29th February 2012
Old 29th February 2012
  #4
Lives for gear
 
popmann's Avatar
 
Joined: Nov 2004
Location: Nashville
Posts: 4,690

popmann is offline
Well if we're going to use Reaper as the benchmark, it sounds exactly the same.
__________________
New singles available in all it's linear 88.2 glory:
Cogs In the Machine
Cliche
Older archives:
My work--now available in linear PCM for download!
geartommy
Thread Starter
#5
1st March 2012
Old 1st March 2012
  #5
Gear interested
 
Joined: Apr 2011
Posts: 22

Thread Starter
geartommy is offline
Quote:
Originally Posted by timlloyd View Post
Nonsense

Luckily for us all, the chaps at Cockos have catered for this very scenario. Download the trial of reaper if you don't already own it.

Do a mix and render it. Then open project settings, click on the "Advanced" tab and change the "Track mixing bit depth" to 32 bit float or whatever takes your fancy, and render again.

Then abx. Then put it back to 64 bit float (cz ... why not), and forget all about it
OK, I did a test:

Reaper 4 under Windows 32-bit
Project in 24-bit 44K
30 tracks in 24-bit 44K
Basic mixing, level adjustments only, no plugs
All tracks summed to the master track
Render the master mix to a 24-bit Wave PCM, no dithering, no noise shaping

And the results (note that I use the word "summing" instead of Reaper's "track mixing depth"):

64-bit floating-point summing (the benchmark)

Vs 32-bit floating-point summing: the files nulled down to -132dB, deterioration was inaudible to me at maximum output levels

Vs 24-bit integer summing: the files nulled down to -115dB, deterioration was inaudible to me at maximum output levels

Vs 16-bit integer summing: the files cancelled each other down to about -64dB, but degradation of the signal below that threshold was very pronounced -- lots of audible noise and harshness

Conclusion:

Obviously, the longer the number, the closer to the original. However, the "noise" that is supposedly introduced by the "sloppier" summing is inaudible to me even with 24-bit integer summing. Certainly, would never make a difference between a 64-bit summed file and a 32-bit summed file. It was surprising how badly the 16-bit summing sounded though. Even though all summing was rendered to 24-bit files, the 16-bit summed 24-bit file itself sounded noticeably harsh and flat. I guess anyone would hear the degradation, even a casual listener if his attention is directed to the noise and lack of depth.

Can you guys get more dramatic results when summing 32-bit, as compared to 64-bit? Unfortunately, I can't compare objectively between DAWs, as replicating a mix in two or more applications would be quite cumbersome. Thankfully, Reaper allowed some flexibility so I could test different scenarios. I don't think Logic allows me to change the math though. Any ideas?!

Would be really happy if someone could try mixing down the same project on different Pro Tools versions, especially version 10. Would be really interesting if someone can get *audibly* better results with 64-bit summing/32-bit plugin processing.

Quote:
Originally Posted by Gear Geek
It is most relevant in plug-in processing and only has and advantage with plug-ins that are using 64 bit processing.
Which plugs would you suggest we test?! What kind of complex processing should take place so that you can actually HEAR a difference when the final mix gets summed 32-bit as opposed to 64-bit?
#6
1st March 2012
Old 1st March 2012
  #6
Gear Head
 
Janne19691's Avatar
 
Joined: Aug 2011
Location: Helsinki
Posts: 63

Janne19691 is offline
If this is the '64-bit double precision engine' option in Sonar X1 then yes i can clearly hear the difference. I have it always on. Still using 32-bit Sonar and plugins. Audio 24-bit 44,1 kHz.
#7
4th June 2012
Old 4th June 2012
  #7
Lives for gear
 
gizeh12's Avatar
 
Joined: Nov 2005
Location: United Kingdom
Posts: 1,103

gizeh12 is offline
So the 64bit summing mixer sounds significantly better?
#8
4th June 2012
Old 4th June 2012
  #8
Lives for gear
 
Joined: Jun 2009
Posts: 1,801

Jon Hodgson is online now
Summing at 64 bit floats versus 32 is not going to make any audible difference unless you're doing something very odd with the result. Geartommy's results are what you would expect, the rounding errors with 32 bit gain and sum are already well below anything that will be audible.

However, the DAW manufacturers don't tell you what they include in their "summing engine", if they actually mean their whole basic mixer, including EQs and filters, then processing those at 64 bits may make an audible difference with certain settings.
#9
4th June 2012
Old 4th June 2012
  #9
Gear Guru
 
theblue1's Avatar
 
Joined: Mar 2005
Location: Long Beach, CA
Posts: 21,158

theblue1 is offline
Quote:
Originally Posted by Janne19691 View Post
If this is the '64-bit double precision engine' option in Sonar X1 then yes i can clearly hear the difference. I have it always on. Still using 32-bit Sonar and plugins. Audio 24-bit 44,1 kHz.
One of the nice things about Sonar's implementation of switchable 32 float and 64 float summing is that you can have it use 32 bit for tracking and monitoring but then automatically use the 64 bit engine for actual rendering. Best of both worlds.

That said, when I first got Sonar 4 (I think it was) with a 64 bit option, I did some informal comparisons and didn't find any casually noticeable difference in rendering between 32 float and 64 float on the (rather small-ish) mixes I tested it on.

But since I can use 64 bit for rendering only, I just do that, since my computer is pretty modest. No tricky decisions, weighing of operating load vs quality, etc.
#10
4th June 2012
Old 4th June 2012
  #10
Lives for gear
 
TranscendingM's Avatar
 
Joined: Jul 2006
Location: NY
Posts: 2,041

Send a message via Skype™ to TranscendingM
TranscendingM is offline
Yea those extra bits are "parked" more or less. If you are just summing there's no difference really. It's in processing where big word lengths come into play.
#11
4th June 2012
Old 4th June 2012
  #11
Lives for gear
 
Dpro's Avatar
 
Joined: Mar 2012
Location: Los Feliz/Hollywood
Posts: 1,025

Dpro is offline
Am I missing something here? Logic works in 64bit now. It free's up a lot of the limitations of working in 32bit. One would assume since it is 64bit cable you can sum in it at 64 bit.
I am only currently using Logic at 32 bit because the particular mac I am using is still just running leopard and not Snow Leopard or Lion. Plus it does not have enough ram in it to exploit the crazy enhanced ram capabilities of 64bit Logic i.e you can actually address 12-16gb's of ram with it VS the 4g limit in 32 bit.
I have heard once you go 64 bit in Logic there is no turning back because of the advantages. As soon as I get my new Mac I will be going 64 bit.
#12
4th June 2012
Old 4th June 2012
  #12
Lives for gear
 
gizeh12's Avatar
 
Joined: Nov 2005
Location: United Kingdom
Posts: 1,103

gizeh12 is offline
Let's make this clear since there is so much confusion over this. 64 bit summing is not the same as 64 bit processing.
__________________
dIal3ctical mater1ali5m
Quote
1
#13
4th June 2012
Old 4th June 2012
  #13
Lives for gear
 
valis's Avatar
 
Joined: Apr 2006
Location: Northwest USA
Posts: 4,676

valis is offline
Quote:
Originally Posted by gizeh12 View Post
Let's make this clear since there is so much confusion over this. 64 bit summing is not the same as 64 bit processing.
No what you mean is that 64bit summing is not the same as 64bits of memory addressing, which is what happens when you switch Logic to 64bit mode.

Ie, 64bit "summing" or internal processing in a plugin uses DATA VALUES that are 64bit FLT (note that data values can use even longer wordlengths if using SSE4/x87 modes that can exceed 80bits of width).

64bit MEMORY ADDRESSING means that an application can access more than 3.3GB of ram (theoretical 4GB but in reality most 32bit apps are limited to either 2GB on Win32 or 3.2-3.3GB on OSX or under 32bit windows with the /3GB boot switch). Thus Logic in 64bit mode can access tremendously more ram for samplers, romplers & VI plugins but Logic *still uses 32bit FLT summing*.

Which answers Dpro's question.
Quote
1
#14
4th June 2012
Old 4th June 2012
  #14
Lives for gear
 
valis's Avatar
 
Joined: Apr 2006
Location: Northwest USA
Posts: 4,676

valis is offline
Now as for 64bit SUMMING, it's nice to access a noisefloor that goes beyond -144dB when using 32bit FLT, but the real reason that mixing engines run in floating point instead of integer isn't just due to the noisefloor issues with quantizating (quantization noise added ot the signal).

First the dynamic range of 32bit FLT is 1530 dBf so you don't have to worry about a signal overflowing (or underflowing) the range of acceptable values in any real world situation--and so you sidestep a bunch of extra work that is done in the fixed point realm to insure that your signals are appropriately scaled (which is where a lot of the added quantize noise in fixed point processing comes from).

Secondly, computers are basically great at handling floating point math, and both the coding languages and modern (Intel/AMD) processors offer extended functions to handle certain floating point transforms very quickly (which is what SSE2 through SSE4 instructions were all about, and the modern AES etc extensions are applying similar shortcuts to algorithms used in encryption/decryption now as well).

Also it's worth noting that the *noisefloor* itself is usually a function of the signal being handled. A signal recorded with a Microphone and a ADC that only has 18bits of precision before thermal noisefloor is still going to have the same self-noise regardless of whether it's mixed in 32bit FLT or 64bit FLT summing, as any quantization error is going to be so far below the noisefloor of the signal's own noisefloor as to be insignificant. And a VI that outputs oversampled synthesized material is still going to have its own internal "errors" and thus a built in noisefloor. But what we're concerned with here isn't just self-noise, but what is called quantization noise...or the errors that accumulate on a signal when processing occurs. So the question is what benefit does 64bit FLT summing offer over 32bit FLT?

Take a quick segue and consider what sort of processing you do in a DAW...in short form you take some signal (from a prerecorded source or from a virtual instrument etc) and you pass it through a series of plugins (that do a variety of processes) and finally 'sum' the signal in the summing engine of the DAW: The most destructive processing will come from plugins, especially ones that are the NONLINEAR by nature, followed by plugins that are more linear but still create enough of a change that they could potentially benefit from an increase in precision (possibly because they're repeating something hundreds of times as with a reverb trail).

In both of the cases here (plugins that are completely nonlinear or plugins that are relatively linear but highly recursive) the increased processing precision can be accounted for entirely within the plugin! So the question becomes, how much quantization noise occurs when you take a signal from 64bit FLT (or higher precision) and quantize it down to 32bit FLT? In most cases that I'm aware of the quantization noise should still be under -300dBf (which is why I wonder about the result of geartommy's test) and so even with a large number of channels you're unlikely to 'add' enough of this 'error' to make an audible impact.

There is 1 other case where the increased precision of 64bit FLT will matter, and that's when dealing with *extremely* large or small values, as the floating point numbers are scaled in such a way that the 'ends' of the scale are much less accurate than the center of the scale. So 64bit precision will result in less quantization noise if you're for some reason scaling values way way way up & down (1000dB!) at the SAME TIME and mixing them together (so that you aren't just manipulating the exponent but combining the result)... But this is not something that is normally a concern in the audio world as far as I'm aware.

Sorry this is so verbose, I'm trying to encapsulate a great amount of material and glossing over a lot here, and while this is probably elementary enough to someone like Jon that he'll be able to parse it quickly and even point out my errors, this is probably tl;dnr to many of you.
Quote
1
#15
4th June 2012
Old 4th June 2012
  #15
Lives for gear
 
Joined: Jun 2009
Posts: 1,801

Jon Hodgson is online now
Quote:
Originally Posted by valis View Post
So the question becomes, how much quantization noise occurs when you take a signal from 64bit FLT (or higher precision) and quantize it down to 32bit FLT? In most cases that I'm aware of the quantization noise should still be under -300dBf (which is why I wonder about the result of geartommy's test) and so even with a large number of channels you're unlikely to 'add' enough of this 'error' to make an audible impact.
Although I agree with your final conclusion, your maths are a little off here.

Assuming we're using rounding quantizing to any bit depth gives +/- 0.5lsb

In the case of 32 bit we have in effect a 24 bit mantissa (there are 23 bits stored, the other is always 1), which means that we're looking at an error in the region of 140dB down on our given signal at any time.

To get -300 dBfs of noise you'd need a signal peaking at -160dbFs, which would be silent anyway.

Looks to me like you're looking confusing the rounding error of a 64 bit float with that of a 32 bit, since the error in a 64bit float is around -312dB below the signal
#16
4th June 2012
Old 4th June 2012
  #16
Lives for gear
 
Joined: Jun 2009
Posts: 1,801

Jon Hodgson is online now
Quote:
Originally Posted by valis View Post
There is 1 other case where the increased precision of 64bit FLT will matter, and that's when dealing with *extremely* large or small values, as the floating point numbers are scaled in such a way that the 'ends' of the scale are much less accurate than the center of the scale. So 64bit precision will result in less quantization noise if you're for some reason scaling values way way way up & down (1000dB!) at the SAME TIME and mixing them together (so that you aren't just manipulating the exponent but combining the result)... But this is not something that is normally a concern in the audio world as far as I'm aware.
Actually it can be, if you get it wrong filters can become unstable, because you can have very small values in feedback loops, if they get rounded away then the feedback doesn't work properly.

It doesn't always require more bits to fix though, changing the order of operations can sometimes be enough.

But really that's a seperate issue from how wide the audio paths are passing data around the system, if the developer of the filter needs to use 64 bit intermediate values then he can do so, regardless of whether the input and output are 32 or 64 bits.
#17
5th June 2012
Old 5th June 2012
  #17
Lives for gear
 
valis's Avatar
 
Joined: Apr 2006
Location: Northwest USA
Posts: 4,676

valis is offline
Quote:
Originally Posted by Jon Hodgson View Post
Actually it can be, if you get it wrong filters can become unstable, because you can have very small values in feedback loops, if they get rounded away then the feedback doesn't work properly
Sorry I must have been unclear, I meant that you to carry the result back to the mix engine in 64bit FLT just to gain the advantages you gain in 64bit FLT when processing.

I knew you'd be able to correct my math too I don't do dsp coding for a living so you'll have to pardon me, just played around with things here & there. Also very tired atm from a show so I'll read the rest later and respond if I missed something else...
#18
5th June 2012
Old 5th June 2012
  #18
70% Coffee, 30% Beer
 
Doc Mixwell's Avatar
 
Joined: Dec 2006
Location: Quincy, MA
Posts: 10,280

Send a message via Skype™ to Doc Mixwell
Doc Mixwell is offline
Not a math geek, but

I find, NOTHING AT ALL WRONG with Digital Audio Summing, these days. I am getting a loootta work done with my Pro Tools rig, and everyone is very happy with my mixes. Mostly, I find people abusing their tools, and then blaming them in the same sitting.
Quote
1
New Reply Submit Thread to Facebook Facebook  Submit Thread to Twitter Twitter  Submit Thread to LinkedIn LinkedIn  Submit Thread to Google+ Google+ 
 
Thread Tools
Search this Thread
Search this Thread:

Advanced Search
Similar Threads
Thread
Thread Starter / Forum
Replies
Agno / So much gear, so little time!
49
day-go / Low End Theory
86
bunnerabb / So much gear, so little time!
0
mahasandi / Music Computers
0

Forum Jump

SEO by vBSEO ©2011, Crawlability, Inc.