Truncation: Inaudible or Significant ???

[djhead]

Does anyone besides them know what language they are speaking?

 

[masteringhouse]

Does anyone besides them know what language they are speaking?

It makes my head hurt too ...

 

[SouthSIDE Glen]

When he discusses the pros and cons of floating, essentially he states that it's easier to program, whereas there seems to be some risk, if not quantified, that fixed could be screwed up by the programmer. Certainly that would have to be weighed by a developer, along with issues such as heat dissipation. However, for a DAW software company, the physical problems of floating-point design have already been solved for them by Intel, AMD, et al. Therefore they simply have a set of specifications for CPU power they must consider.

As a developer, I'd like to throw in my couple of least significant bits on this one

The hardware is only the second half of the equation. What matters to the developer - more accurately, the project architect - is the capabilities of the programming language used, as it is the programming language that talks to the hardware and instructs it how to channel the bits. Without getting too technical, where the software needs to be architected correctly is in when and where it uses different variable types (e.g. long integer, short integer, floating point, etc.) not to mention the proper translations when pushing bits thourgh assembly language libraries, as well as other architectural considerations. It's not quite as simple as "just make everything floating point"; there are many factors involved including memory allocation management, speed and efficiency of individual code modules, thread management, and other technical mumbo jumbo.

Is it easier on the coder to just make it all floating point? In some ways, yes. Is that always the best way to actually design the program? Not necessarily.

G.

 

[mshilarious]

MS -

I don't understand the relevance of power consumption or heat dissipation. The point made by the article is in the way floating point and fixed perform calculations.

I believe it is also a trivial issue, but it was brought up in the article as an advantage to fixed--simpler calculations mean fewer transistors which means less heat. In Rane's world of rack gear, where a chip is selected for DSP, that has relevance, but in PC DAW-land the transistors are already there, and maybe even overclocked

In regards to using float as a crutch, I believe the point there was that some budding engineers do not pay proper attention to gain staging and rely on 32 float processing to cover overs rather than making the proper adjustments to fit within the contraints of 24 bit. If you consider that there may be some additional inaccuracy when processing beyond 24 bits in a 32 bit floating point register, it may have some theoretical merit.

That is what I was getting at--the article's point seems to be that float does let you get away with sloppiness in gain staging, at the cost of errors in summing, irrespective of gain stage. Which is more useful? For me, if I can demonstrate that the effect of the summing error is random, and lower level than dither, I cannot see or hear any harm.

And I freely admit to somewhat sloppy gain-staging. I will let the level float between plugs wherever it goes. Generally when I am done, I'll fix levels all the way through, but when I am throwing in plugs at various tracks and experimenting, it's a big time saver. And I know that from testing, doing +24dBFS mixes and then chopping the master fader nulls



You know, my previous DAW was a 32 bit fixed engine; I really can't tell any difference. Of course back then I would bring in tracks that peaked at -20dBFS through 18 bit converters, and then normalize them Shhh, don't tell anybody

All this gear I have now is so much better than anything I could have dreamed back in 1997 when all I had was a Tascam . . . really, if I can't produce a decent track with 117dB dynamic range converters and a 32 bit mix engine, fixed or floating, then I pretty much suck . . .

 

[SouthSIDE Glen]

really, if I can't produce a decent track with 117dB dynamic range converters and a 32 bit mix engine, fixed or floating, then I pretty much suck . . .

That's probably more self-depricating than you need to be .

I agree, though, that this topic belongs in the same folder as the 48k vs. 96k topic, perhaps even more so. If one can't get a pleasant-sounding production out of fixed point, floating point isn't going to help it sound any better.

G.

 

[LeeRosario]

MS -

I don't understand the relevance of power consumption or heat dissipation. The point made by the article is in the way floating point and fixed perform calculations.

In summary the point that I believe he is making is that floating point inherently has 24 bits worth of precision no matter what range of audio you are trying to represent, while fixed point calculations are usually performed at a higher 48 bit precision and then brought back to 24 bit.

The design of the application is key. Some fixed plugs will dither back to 24 bits after calculations in 48 bit, others don't. It's my opinion that those that truncate sound colder when doing this. IMHO all fixed plugs that work this way should at least have an option to turn on dithering, Waves plugs do this. It's also the reason that I use the dithered mixer in a Pro Tools TDM rig.

In the case of floating point calcs, the article sugggests that there are innacuracies once the precision does beyond the 24th bit and can result in a pumping noise floor. Check out the Moorer article for more info on this.



Is it audible? I'll let others decide that.

In regards to using float as a crutch, I believe the point there was that some budding engineers do not pay proper attention to gain staging and rely on 32 float processing to cover overs rather than making the proper adjustments to fit within the contraints of 24 bit. If you consider that there may be some additional inaccuracy when processing beyond 24 bits in a 32 bit floating point register, it may have some theoretical merit.

Delicious advice for the masses

 

[masteringhouse]

Delicious advice for the masses

Thanks Lee, though I think at least some of this conversation bored the average reader and was ignored about 5 posts ago.

 

[mark4man]

...though I think at least some of this conversation bored the average reader and was ignored about 5 posts ago.

Not by this average reader.

I'm reading, learning, enjoying & appreciating the input.

mark4man

 

[mark4man]

Here's what I'm wondering in all of this:

Are these inaccuracies associated w/ floating-point architectures significant enough to cause one to switch over to a fixed-integer system (say, Pro Tools) from a floater (say, SONAR)?

In other words...can the inaccuracies be heard in the audible band?

mark4man

 

[SouthSIDE Glen]

Personally, my answer is more towards the line of, "If it's truely time for me to be seriously worried about the differences between fixed and floating point software, I am a million miles ahead of the game for having all the far, far more important stuff down cold."

Stuff that will have a far more important effect on my sound like "how to get the perfect tracking" and "how to get the perfect mix" down cold, that I'm doing that with microphones, preamps and converters that give me at least one channel into my DAW actually worthy of either fixed point or floating point, and that both my tracking room and mixing room are treated to let me actually get the best out of the gear that I do have.

Until then, I'm satisfied that the sound I get out of Steinberg and Sony editors are just fine for what I'm putting into them, and will worry about the esoterics of that final end of the signal chain when it actually makes sense for me to do so.

Just one personal viewpoint.

G.

 

[masteringhouse]

G. is right, the difference between float and fixed is going to be subtle over the things that go into it.

The sonic issues are more critical when you are performing multiple stages of processing (like in mastering) and the effect becomes additive. In general as long as your plugs process at a greater bit depth internally and dither back to the target depth properly and you adhere to good gain staging practices you should be in very good shape with either.

BTW Pro Tools does both float and fixed. TDM systems are fixed, LE systems are float. Additionally I believe RTAS on a TDM system is float since it uses the host system for processing.