I need to add one further post because, based on reading a thread with Paul Isaacs, my latest understanding is that the Sound Devices new Mixpres operate in the exact same way as the Zoom F6. In both brands, 32-bit float mode uses the gain knobs purely as post-converter digital gain faders (i.e. they introduce zero noise). In other words, you cannot set traditional gain levels in this mode. The only difference is that the Zoom uses 2 preamp/convertors combined in a clever way and the SoundDevices, 3. So, my earlier thinking that I would set a healthy approximate gain on sound check is completely unnecessary (and impossible) other than to get a decent monitoring level. You really can leave the “gain” knob at its lowest setting and bring it up in post with no downsides. Someone jump in and correct me if I’m wrong!
If it is true that any gain/level control is post digital conversion and post conversion to float, then that is correct. There would be no sound reason for setting gain before converting to float so if there is no possibility of setting analog gain then the record level is not relevant. In fact the level control may not even affect the output digital level at all only the monitor level. I think for such a method of conversion to work right, the input level may require a factory setting or auto setting… though that doesn’t sound right either. The mic type would make a difference too. There is a difference between dynamic, condenser and ribbon mics for sure. Maybe there is a switch instead of a continuous level for that though.
This begs the question: If some sort of auto level is included, does that reduce the dynamics of the audio being recorded? Or is there a setup before recording to get it right before recording so the level(s) can be locked? In theory, it should be possible to get a better setting by doing this step manually but even more complex than setting up a normal single conversion level. If there is no auto analog level setting, then the mic used matters (note that the zoom comes with it’s own mic, at least the ones I have seen). Of course with 24bits of resolution (for 32 bit float), which is probably 8 more than needed, the extra 48dB of range may cover this quite well. The three converters may be as much as 20dB apart. The level range from my ribbon to my condenser mic is 25 to 30 dB and would fit into that range with no problem.
I believe that the level control will affect how the 32-bit float file is printed or at least that’s how both the SoundDevices and Zoom are presented when showing recordings at lowest, middle and highest “gain” settings. i.e. moving the knob all the way to the right will print a file with the waveforms looking like they are clipping given enough input. Of course, the beauty of the float file is that reducing the gain in the DAW will reveal the unaffected peaks.
There’s no “auto” level, simply a set gain structure that has a wider dynamic range than any currently-made microphone. All dynamics would be available.
Not true on the F-series models. The F6, currently the only model that includes 32-bit float recording, has 6 xlr connections. The 32-bit stuff is only dependent on the preamp/converter stage.
My understanding is that the unknown “magic” is the way the companies glue together the final 24 bits (23-bits + sign) with the 8 bits of exponent for the digital gain by taking, say, 12 bits from one preamp/converter level and 11 bits from another (Zoom) and a three-way equivalent, say 7+8+8 for the SoundDevices. I suppose it isn’t that simple but that’s how I’ve seen other people talk about it. Perhaps the distribution of bits is not as equally distributed, for example.
EDIT: I should also add that it really is impossible to clip the internal circuitry of both the Zoom and SoundDevices as the uppermost limit of the gain structure aligns with (or is a hair under) that clipping point.
That does not make sense to me. In my mind there would be three ADCs running in parallel, each with a different gain. ADC 1 would be set at 0 dB, ADC 2 would be -20 dB and ADC 3 would be set at -40 dB (for some suitable value of 0dB). When quiet sounds are encountered, ADC 1 would have the greatest number of bits representing the audio and so it’s output would be used and converted to 32bit float and then it’s gain reduced by 20dB. When the input got loud enough that ADC 1 was over loaded then ADC 2 would be used converted to float with no gain change. Then when the input got even louder and ADC 2 is now clipped, the input from ADC 3 would be used and of course after being converted to float the level would be increased by 20 dB.
This setup would require that the 20 and 40 dB pads in the analog section were precise (do-able) and that the ADCs used are matched (all 6 of them for stereo). Also note that the ADC output would be chosen on a per sample basis so that even on loud sounds, a sample that is closer to a “0” value (not talking dB here but rather 24bit int) would still use ADC 1. It would be much easier to explain with a drawing, but I am not that kind of artist 
Please also note that my choices of dB steps is only for illustration purposes and may be wildly different from the reality. I do expect they would use “even” amounts of gain so that the 24bit part of the float could remain the same and only the exponent would need to be changed… well actually it would not be changed but rather a specific exponent would be added to the output of each ADC when converting it to float so that there would be no math needed. Switching would be achieved by simple “less than” “greater than” which could probably be done with simple register compare rather than real math. Not magic really, but more than 3 times more costly. I do not know if the gain is really worth the cost… yes you could use real canons … but in the end the output will be restricted to 16bits or maybe 24bits and the speakers would need to have the available power output equal to said canons. Hardly front room listening levels.
Not having to set levels in a high paced film recording situation would be another use. As for the actual sound, one would have to do double blind AB testing. Better if you have one to just enjoy…
My head hurts. Diagram, please 
In all seriousness, yes I plan to just enjoy it. But my brain does like to figure out what is going on. Paul Isaacs refuses to go into details other than about the “gain” being digital post-preamp/ADC.
Nah, that is the canons going off…
If someone already linked to this earlier, sorry, but here’s a link to Sound Devices’ patent application for this technology: https://patents.google.com/patent/US9654134B2/en
This is great, thanks @bradhurley!
So, while I don’t understand most of those diagrams, am I seeing three separate pre-amps and Analog-to-digital convertors? That would align with how the Zoom F6 has been described albeit with two pre-amps and ADCs. What I don’t get is where it says “analog gain/filter 1”. It does seem at first glance that we are not dealing with pads as @lenovens suggested but I await his rebuttal with tingling anticipation. Also looks like there is a buffer involved to allow processing time to figure out best fit of the bits…
I noticed that they were aimed directly at your moon real estate.
No rebuttal, I would consider gain stage to be equivalent to pad for the purposes of this discussion. Gain is a ratio of output over input. A pad has a gain of less than 1.0 and an amplifier has a gain of greater than 1.0. A buffer generally has a gain (voltage gain) of 1.0 while offering greater current. I chose to suspect pads because they are easier to get accurate being a resistive network only. The patent as presented does not preclude the use of pads but rather includes them as well as amplifiers. The example provided only seems to extend the resolution by 22dB rather than the 40dB I suggested (which is why I said my values could be wildly off). However, the patent is only an example of the idea and does not need to exactly represent the actual values used inside. 22 seems odd too, I would think 24 would make more sense as 12dB is a gain of 4 (I think), so two gain steps for three ADCs would be 24dB. no matter what values they actually use, this patent would cover them.
Other than that, the patent is pretty close to what I suggested off the top of my head. It is a good idea… but it would have made more realistic sense 40 years ago when 16bits was hard to get and 24bit converters were not around. The difference would be much more audible… or at least it would have had a chance of being audible. At this point it is more of a convenience rather than an improvement. Still worth while for all that.
I’m disappointed. So a “buffer” which to my previous understanding simply meant a period of time to store data for processing now “has a gain of 1.0 while offering greater current”? If this be true, I’m clearly in over my head. Still got that brochure on lunar properties?
Roughly, yes.
12dB ≈ 3.98…
The formula is :
dB = 20*log10(gain)
so that only comes out as an integer when the gain is a power of 10. 20dB = 10, 40dB = 100
Turning the formula around gives :
gain = 10^(dB/20)
so for 12dB
gain = 10^(12/20) ≈ 3.98107170553…
an unwieldy decimal conveniently rounded to 4.
There are buffers and buffers. In the digital world, a buffer is a storage area and generally causes delay, in the fire fighting world, a buffer is a bare area, in the woman’s monthly world we use the french word for the same thing… In the analog area (which we were talking about) a buffer is something that keeps an input from overloading an output in such a way that the output audio (or other signal) is not what is expected. So in this case a buffer is an amplifier that does not add or subtract gain but does allow the next part of the audio chain to draw more current. All amplifiers tend to do some buffering even if they also change the gain but they called a buffer if their function is to only provide a higher current output.
Brochure not needed, just send money
So, just to be clear, in this instance we are not talking at all about a buffer as a period of time while the device figures out which bits to use? i.e. a delay in writing the audio to the SD card to allow for calculations?
In this case the gain is more important (voltage gain not power). We talk about dB gain because it makes more sense to our ears but in this case the actual gain is more important because the input number is is a 24bit int which we don’t want to manipulate. rather we want to hard set the exponent part of the float we create to give the gain we want. If that amount of gain is an odd amount, it doesn’t matter. It doesn’t even matter if the gain steps are not equal. The numbers just have to fit. I do not know my binary math well enough to give reasonable examples, I chose 4 as an example because it is an even binary number… but the exponent part of the float would be chosen not to be exactly a gain of 4 (unless it just happened to work out nice) but rather a value that allowed the significand to remain the same. In other words it is important for the gain steps to translate to “nice” numbers rather than easy for humans to read numbers. So after figuring out the gain, the closest whole number in dB would be stated… or perhaps the number that rolls off the tougue the best as is the case with engines (what size really is a Chev 350?).
A buffer is never a period of time even in the digital world though it may have the effect of creating a delay. In the digital world a buffer is a place to keep things until we are ready for them. A place in memory to put stuff if you like. The reason such a buffer is needed is that even if we were dealing with one sample at a time, there are two things accessing this buffer and we do not want (ever) both things to try to access this buffer at the same time. So in the case of audio, the ADC writes to a buffer some amount of samples (1 or more but normally at least 16). When it has finished writing to that buffer, it sends a signal to the computer (an “interrupt”) and the audio software in the computer then knows it is allowed to read that buffer and process the audio it holds.
And no, we are not talking about that kind of buffer. In this case I was talking about an analog buffer who’s job is to match one amplification stage to another. These are generally high impedance input and low impedance output (high current output) devices.
gain = 10^(12/20) ≈ 3.98107170553…
That is the voltage gain I posted. The formula to convert power gain in dB to a number is 10^(dB/10).
I am sorry, I did not mean to indicate you were wrong so much as make sure others reading knew what we were talking about. Or at least knew they are different.
Meanwhile, I’m having lots of fun with my Mixpre6 mark ii. Just recorded concert with the knobs all the way down just because. Glad to know we have the numbers covered
No need to apologise. I too only wanted to clarify.
When you’re taliking about ‘32 bit float’ are you referring to IEEE 754 with sign, exponent and fraction like this?
s|eeeeeeee|fffffffffffffffffffffff
That is put back together as +/-(1+fraction)*exponent. The exponent takes the value of 2^(e-127). The gain then would take power of 2 values … 1/8,1/4,1/2,1,2,4,8 …
If the exponent was to be used as a gain control it can only work in 6dB steps.
I may be missing something.
Nope you have it exactly right. Three ADCs. Use the 24 bit ADC value for the sign and fraction. The exponent would set the ADCs output to:
ADC1 would be either +6, +12 or +18 (it seems +12 is what is used)
ADC2 would be set to 0
ADC3 would be set to -6, -12 or -18 (it seems -12 is what is used)
This means the “gain stage” in front of each ADC needs to be precise. The ADC with +12 in the conversion needs to have a gain of exactly -12 and so on… well it would not (as you have already pointed out) be exactly 12 dB but rather the ADC conversion would be a gain of 4 and the gain stage would be a gain of 1/4. This is why I suspect that a pad is used for the “gain stage”. So:
Mic -> pre -> split -> (-24) -> ADC-24bit int -> (+24) 32bit float -> switch -> output
|
-> (-12) -> ADC-24bit int -> (+12) 32bit float ^
|
-> (0) -> ADC-24bit int -> (direct) 32 bit float ^
Not sure how my diagram will turn out… Anyway the switch selects between the three outputs from the bottom up. So long as the bottom ADC’s 24 bit value is between 1 and -1 that value is used, If the value is either 1 or -1 then the next up value is looked at and so long as it’s value is between 1 and -1 then it’s value is used but if it’s value is also 1 or -1 then the top ADC’s output is used. In my mind, the “gain stages” are all pads (well the bottom one has none) because they would then be simple resistive networks that tend to wander less with time. They are probably situated where the temperature is reasonably constant and at least one of the resistive components would be variable set at factory for exactly the right gain.