Originally Posted By: Gordon Scott
The weakest links in the chain with floating-point like this are the pre-amps+ADC at the front and the DAC+PA+Speakers at the back, and the article glosses over that a bit.. Both of those still need attention to getting the levels sensible and getting the signal into its initial digital format (invariably, I think, non-floating). A quick look at Cirrus and Burr Brown ADCs shows none better that 127dB SNR. Even achieving that with a mic pre-amp as well is a real challenge (e.g., Focusrite show ~110dB dynamic range on their 18i20). It's possible to do better if money is no object.

So whether integer, fixed point, or floating point, one definitely still needs to set the front-end gain/level sensibly.


Once into the digital domain and converted to floating point, one can thenceforth essentially ignore signal levels and gain, right up to the point where she signal goes back to non-floating-point, or to analogue at the amplifiers/speakers.

The significant-digits part of 32bit floating-point is 24-bits, so exactly the same resolution/precision as 24-bit integer data, but it adds also an 8-bit exponent (10^+/-127), which is huge for audio purposes.

If one changes the level many times, by arbitrary amounts, the signal will eventually degrade, due to rounding errors in the calculations, but in practice it's by such a small amount that it's likely irrelevant. I'm not confident of the maths here, but I suspect that would be around 32,000 manipulations to degrade the signal to the equivalent of 16-bit CD quality. I can live with that smile


I'm expecting the next generation of inexpensive interfaces to use a chip that supports a 384MHz sampling rate with 32bit float. Although such units exist already (some with 64 bit float), they're pretty expensive. When that happens, just watch: 32 bit float will be sold to us as the only thing we can use. As you've pointed out, it depends.

The current inexpensive chip is 192/24 (aka Gen 3 since so many are using it in their 3rd generation units) and found in 3 dozen or more current interfaces. The eight I/O version with two 'phone amps costs $6.73 ea with a minimum quantity of 1,000. The 4th Gen doesn't have to be that inexpensive—if it can be delivered under $20 per, that should suffice.

Going 32 bit float in my DAW increased my per-track headroom noticeably but I'm putting together projects with 50-100+ tracks of vocals recorded in home studios. I need all the help I can get. Not doing 64 bit float, however, though I could—even Logic Pro X supports that. Certainly don't need anything greater than 24 bit bwav when I bounce.

BIAB 2024 Audiophile Mac
24Core/60CoreGPU M2 MacStudioUltra/8TB/192GB Sequoia, M1 MBAir, 2012 MBP
Digital Performer11, LogicPro, Finale27/Dorico/Encore/SmartScorePro64/Notion6 /Overture5