32 bit DAC?

[stelleg151]

Just wondering if there are any 32 bit DACs out there, as windows Vista will be doing all processing in 32 bit form, and potentially outputting 32 bit audio, so it would be cool to keep it in 32 bit form while doing the conversion.

Any DACs that do this? Is the conversion from 32bit to 24bit anything to worry about?

[Petter]

As far as I have been able to find, nobody manufactures DAC's capable of handling more than 24 bits.

The key to understanding what is going on with Vista is that Microsoft have moved to 32 bit float for the audio path. They can handle 384KHz at 32 bits through the entire system withtout needing special drivers etc courtesy of the UAA driver model.

Now, a 32 bit float is not the same as a 24 bit integer and there are some problems with a 32 bit float. The anatomy of the float is such that you have 23 bit mantissa and 8 bit exponent (and the exponent is shifted so you can represent small and large multipliers), and since this is audio data, there is a sign bit. It should be possible to use the 23 bit mantissa + the sign bit to represent 24 bit integer data accurately but I don't know how this is actually implemented.

A great upside of the floating point format is that we can now do all sorts of calculations on the audio data without the same fear of insertion of wrong data due to rounding errors. Also, as far as I have been able to find out most CPU's handle 32 bit float calculations with 64 bits of precision. The format change will be very useful for room correction, digital crossovers etc, and it can be done on the CPU both in the producing/mixing stage and the playback chain. A musican friend of mine claims 32 bit float mixing work produces much better results than 24 bit, and from studying the spec over the past several minutes I understand why. You don't get rounding errors everywhere, and you can store a rounded number.

Keep in mind that 24 bits is very high resolution and that true 24 bits will likely be better performing than every single following item in your system (at normal atmospheric pressure you need a nuclear bomb to be able to get close to 24 bits of resolution if you start at the low limit of human hearing, and that only provides half the waveform). That is not to say that we are at true 24 bit performance yet .... Also, bear in mind that creating true 24 bit audio data from non-synthetic sources is also very hard if it has even been done yet. Having said that, I think the move to 32 bit float is natural and positive. In fact, I would have preferred that the distribution formats for audio moved to a float as well. Research indicates that an optimal system with ca 19 true bits is very close to the limits of what humans can hear. However, have you ever seen an optimal setup? For audio capture and processing though, many more bits would likely be required.

Now back to your question - can we do a 32 bit float conversion to analog? Current converter architectures seem to be integer based for the most part. Perhaps someone with a huge budget will eventually produce a commercial floating point converter. However, if your source data is integer based and you are doing bit accurate transfers, you would likely be better off with an integer converter all things being egual.

Petter

[lohk]

Since a long time, actually almost since the beginning of serious digital audio most computer audio editors or equivalent programs internally use 32bit floating point only - less is simply not acceptable for any kind of transformation. The floats are always precise as possible of course, as a matter of fact.
The output is either converted into a 24bit or 16bit file according to the requirements, or can also be played back immediatly by just changing the digital values - convert the 32bit float to 24bit integer. The rounding error is minimal, normalisation is although sometimes needed (like a float file can be "very quite" for example without getting unprecise like the equivalent integer file).

As somebody who works with digital audio every day I agree with Petter about the file formats: It would be much better and easier to change to 32bit floats as a basic format, I do that with production files very often, because I do not wnat to loose any resolution and quality. But files do get much bigger though...

[Morning_Star]

The headroom for floating 32 bit is downward. Processes are done using 32 bit so that there is no clipping but there is really no need for it other wise. 24 bit will be the new standard for a very, very long time. For more information about this Massive Mastering and others really break it down at futureproducers.com . I can't remember where but it has been covered several times. They really know thier stuff.

[lohk]

And of course there are 32bit float DACs:
(and I forget to mention, that my latest software for audio manipulation internally usues of course 64bit precision in and out...)


Highest-Performance Floating-Point DSP
for High-Quality Audio Systems
TMS320C6713
Get datasheets and app reports at: www.ti.com/sc/device/TMS320C6713
With up to 300-MHz operating speeds and integrated peripherals, the
TMS320C6713 DSP offers unparalled 32-/64-bit floating-point performance
ideal for professional and high-resolution audio applications.
Key Features
• Up to 2400 MIPS/1800 MFLOPS
• Native instruction support for IEEE 754
• 32- and 64-bit data
• L1/L2 cache memory architecture
• 32-bit external memory interface (EMIF)
• Glueless interface to SRAM, EPROM, Flash, SBSRAM and SDRAM
• Two multichannel audio serial ports (McASPs)
• Wide variety of I2S and similar bit-stream formats supported
• Integrated digital audio interface transmitter (DIT) supports:
• S/PDIF, IEC60958-1, AES-3, CP-430 formats
• Up to 16 transmit pins
• Enhanced channel status/user data
• Two I2C interfaces
• 16-bit host port interface (HPI)
Applications
• Professional audio equipment: digital mixers, broadcast studio
equipment, musical instruments
• High-end audio conferencing equipment
• High-end commercial audio systems

[lohk]

Well, I was joking - this is obviously "only" a DSP, as far as I know there is no real way to make a floating point DAC. How can that work? I cannot see any way to do that without changing to integer.

[phn]

I thought the hi-rez crowd would be all over the idea of a 32-bit DAC like a dog in heat. Of course, 48 bits, not to mention 64, have to be so much better. Just imagine how nice a 64-bit signal must look in your oscilloscope!

[stelleg151]

Sweet thanks for the replies guys, especially Petter, very informative. Yes from what I gather really the 32bit is only necessary to avoid clipping when mixing, equalizing, etc. I guess if the source is 24bit int the conversion to 32bit float should be flawless, but the conversion back post-processing will require rounding that will effect the output.

I hope the rounding isnt audible, but it seems if possible it would be nice to avoid it alltogether.

Ill check out that site morning, thanks.

[Ulas]

Originally posted by lohk
as far as I know there is no real way to make a floating point DAC. How can that work? I cannot see any way to do that without changing to integer.

There are “floating-point” DACs. They are called multiplying DACs. See MX7837/MX7847, for example. Or, just use any two DACs: The "exponent DAC" sets the reference for the "mantissa DAC."