spheris
Max Output Level: -84 dBFS
- Total Posts : 350
- Joined: 2/17/2004
- Location: Retired
- Status: offline
32bit vs 32bit float - Settling the debate once and for all
This has been a running battle in-house here for about 3 years now.
Can anyone weight in on the merits of ether of these formats on being the more precise, or economical of processor throughput as it's gone on too long and I'm tired of sending two of my asst engineers to their perspective corners to stare at each other (I'm waiting on one to stick out their tongue so I can suspend both) but I'd rather it not come to that.
I know there is an answer to this one, but it seems the little information on it all only convolutes the question more. SO help IS appreciated
|
BruceEnnis
Max Output Level: -58.5 dBFS
- Total Posts : 1665
- Joined: 11/5/2003
- Location: Maryland
- Status: offline
RE: 32bit vs 32bit float - Settling the debate once and for all
July 03, 04 10:42 AM
(permalink)
What does this have to do with SONAR
|
spheris
Max Output Level: -84 dBFS
- Total Posts : 350
- Joined: 2/17/2004
- Location: Retired
- Status: offline
RE: 32bit vs 32bit float - Settling the debate once and for all
July 03, 04 10:51 AM
(permalink)
It has to do with end output, we all know sonar processes at 24bit at output and 64 bit internally, but sonar is not the final stage of production. This thread is to settle a format question concerning post output to a dedicated audio editor for post since cakewalk refuses to include an audio editing facility into the current sonar versions that would make the question a non issue as it could all be done internally then at sonars fixed 64 bit path width and output at whatever the upper limit would be depending on the project target.
|
wogg
Max Output Level: -57 dBFS
- Total Posts : 1819
- Joined: 11/14/2003
- Location: Columbus, OH
- Status: offline
RE: 32bit vs 32bit float - Settling the debate once and for all
July 03, 04 2:14 PM
(permalink)
Have you seen this thread?
|
spheris
Max Output Level: -84 dBFS
- Total Posts : 350
- Joined: 2/17/2004
- Location: Retired
- Status: offline
RE: 32bit vs 32bit float - Settling the debate once and for all
July 03, 04 2:42 PM
(permalink)
many thanks wogg, but that was the thread that started (restarted) the debate. It sheds no light on the question at hand though
< Message edited by spheris -- 7/3/2004 2:50:28 PM >
|
cAPSLOCK
Max Output Level: -69 dBFS
- Total Posts : 1071
- Joined: 11/28/2003
- Location: Dallas, TX
- Status: offline
RE: 32bit vs 32bit float - Settling the debate once and for all
July 03, 04 5:57 PM
(permalink)
I think you mean 24bit fixed versus 32 bit floating.
As far as it goes, the majority of the advantage is in the resolution retained when processes like filters or other frequency domain DSP are used.
So then the most important place to be concerned about floating point is in the PROCESSING as opposed to the file storage.
The question is whether or not your processing tools properly scale the filters if they are staying in the fixed point realm. If they do, then they can produce great results with less processor overhead. If they don't, then you are losing resolution.
This whole thing is one of the reasons... well the main reason why digital filters (EQs) tend to suck.
Now that brings us to the ever looming question of dither, when and how much? This depends entirely on your 'digital signal chain' and when you store the files to a 'lesser' format as a file, or even if the dsp you are using does it's processing at higher than operational bit depth then if you want to be a perfectionist about it, you have to dither... but the point at which you dither when using floating point is complicated can only be determined by the peak amplitude of the file itself since foating point scales your full 24bit resolution. In the case of fixed processing the process is much simpler since you ALWAYS know where your resolution floor runs out.
In the end, as it applies to sonar, files are stored in fixed 16 or 24 bit. So if you are working with sonar it is a mootsih point. Unless of course you capture the output of the audio engine with some sort of real time plugin. Frankly, I don't even know what Sonar's audio engine is outputting at the different points in it's virtual signal chain... does someone know?
Oh, and not only do you have to care about thet if you are going to care at all... but you need to know what every plugin and piece of outboard gear is doing to the bits as well...
Feh. ;)
cAPS
"We da da sahw pe paw fidlily-doobee afidlily-dooten-bweebee!" -Shooby
|
spheris
Max Output Level: -84 dBFS
- Total Posts : 350
- Joined: 2/17/2004
- Location: Retired
- Status: offline
RE: 32bit vs 32bit float - Settling the debate once and for all
July 03, 04 6:03 PM
(permalink)
Almost enough to make you dispair isn't it?
the real question is, if given choice, would floating point 32 or straight integer 32 the better of the two formats to use. Damn sonic foundry and them giving us choices ;)
|
cAPSLOCK
Max Output Level: -69 dBFS
- Total Posts : 1071
- Joined: 11/28/2003
- Location: Dallas, TX
- Status: offline
RE: 32bit vs 32bit float - Settling the debate once and for all
July 04, 04 0:52 PM
(permalink)
***Slightly pedantic reply deleted in favour of the next post. ;)***
< Message edited by cAPSLOCK -- 7/4/2004 1:11:10 AM >
|
cAPSLOCK
Max Output Level: -69 dBFS
- Total Posts : 1071
- Joined: 11/28/2003
- Location: Dallas, TX
- Status: offline
RE: 32bit vs 32bit float - Settling the debate once and for all
July 04, 04 1:09 AM
(permalink)
OK.. I give in... The simple answer to your question is just this: 32bit fixed would be better than 32bit/float for three reasons. 1. 32bit fixed has a higher end resolution than 32bit float. Although 32bit float is distributed over the same data resolution the end product is still a number with a 25bit mantissa and the extra data used to place this within the full range of the dynamic spectrum. This means it is just a virtually perfect 24bit signal instead of a real 32bit signal. 2. Floating point computations lose the double presicion with fixed pont math in most DSPs and CPUS, AND gain the added problems of a possibly shifting noisefloor. 3. Economy as you mention. Floating point math requires much more of a DSP or lowly general purpose CPU. The chip will run hotter and consume more power. This fails to take into account many of the other important considerations like what your D/A converters can actually handle, whether or not you really need protection against 'digital clipping' (which you MAY depending on the application), and finally whether the recording can even come close to utilizing the full 1500 dB resolution afforded to you when you factor in the noisefloor and sensitivity of the gear you are using from the microphones all the wad up the signal chain to the converters, and of course then back out again. Your question is easy assuming you really do mean 32bit float vs 32bit fixed. If you really mean 24bit float versus 32bit fixed (which is a more fair comparison) then the argument will rage on. There are advantages and disadvantages to both formats. cAPS
< Message edited by cAPSLOCK -- 7/4/2004 1:15:25 AM >
"We da da sahw pe paw fidlily-doobee afidlily-dooten-bweebee!" -Shooby
|
spheris
Max Output Level: -84 dBFS
- Total Posts : 350
- Joined: 2/17/2004
- Location: Retired
- Status: offline
RE: 32bit vs 32bit float - Settling the debate once and for all
July 04, 04 2:01 AM
(permalink)
perfect answer and precisely the question
Now we can put this to bed, I did take the time to read the ieee forum paper on it, but it really was never clear on those three points.
caps, I do thank you for the clarity
"Genuine brilliance is a simplified formula - one part egoism, to two parts genius, add a bit of trial and suffering mixed with an optimism towards existentialism..the rest comes with time"
|