192dB is the dynamic range of
Last edited:
-
WANTED: Happy members who like to discuss audio and other topics related to our interest. Desire to learn and share knowledge of science required. There are many reviews of audio hardware and expert members to help answer your questions. Click here to have your audio equipment measured for free!
32 Bit Float Explained
- Thread starter Eurasian
- Start date
May I suggest you read the link that started this thread?
Ah, correction, 32 bit integer has 192dB dynamic range. I doubt this device has the analog frontend and ADC to actually be that sensitive. That would be orders of magnetide better than anything out there.
It’s been tested to have 136dB dynamic range on that video. For some reason this was a negative
While very impressive, still very far of 192 dB and still within the 144 dB 24 bit integer gives us.It’s been tested to have 136dB dynamic range on that video. For some reason this was a negative
If the YouTube guys tests are to be relied on.
Did they? I couldn't find any actual listed specs. In the manual they say that the dynamic range is "amazing" (and it is, compared to most devices out there).
Also, those input sliders are described in the "Adjust the waveform display size" section:
Which hopefully underscores the point that they're not for changing the actual device gain (which is fixed), but act more as a visual convenience. The DAW/app does receive a different input level, but it's all been amplified after the ADC has finished its job, same as if you changed the volume in post.
If they didn't put those sliders there, there wouldn't be a need for the 32bit float format either. But the problem is that people would record a loud signal and it would still be at something like -30dBFS in the DAW (because of the ADC headroom that prevents clipping). Confusion would ensue and Zoom support would be flooded with people who think there's something wrong... So this built in 32bit float "mixer" maybe prevents some confusion and makes users feel good about their gain staging. Still, I would have liked a more honest marketing.
Yes. Here is where the 25 bits and >1500dB come from:
Also explanations about the 25 bits from two SoX developers, robs and Mans Rullgard (mansr in this forum):
For those who are interested in the formats, recently I reported my findings to the WavPack developer (Bryant) about exploiting the relationship between fixed and floating point formats to improve lossless compression efficiency, and a preview was published for evaluation, as well as an updated WavPack Adobe Audition plugin. Bryant also explained how WavPack encodes fixed and floating point files using different approaches.
Improve compression efficiency by treating 32-bit fixed point as float.
Improve compression efficiency by treating 32-bit fixed point as float.
hydrogenaud.io
danadam
Major Contributor
- Joined
- Jan 20, 2017
- Messages
- 1,014
- Likes
- 1,582
"For example, 8-bit audio––the bit depth of choice for the most iconic video games in the 80s and 90s––results in sounds that are flat and robotic since it only contains 256 possible amplitude values"32-Bit Float: Everything You Need to Know
zoomcorp.com
Seriously?
Posted above where they said it.
Confusion will be caused by people who doesn’t understand the concept, which seems to include you from what you said above.
The iconic FM synthesizers used in 80s and 90s video games are based on floating point math too."For example, 8-bit audio––the bit depth of choice for the most iconic video games in the 80s and 90s––results in sounds that are flat and robotic since it only contains 256 possible amplitude values"
Seriously?
Ken Shirriff's blog
Computer history, restoring vintage computers, IC reverse engineering, and whatever
Yamaha OPL - Wikipedia
You're missing the fact that the value is always* normalized, it always has a leading 1. That 1 is not stored, so you have the normalized bit, sign bit, 23 explicit mantissa bits = 25.
*(unless denormalized, for tiny numbers—DSP code always ensure this doesn't happen, because of the performance hit)
OK. The times I used this was years ago for implementing various radar digital signal processing. AFAIK nothing was normalized so we saw 24-bit resolution. That was enough resolution, and FP kept interior (intermediate) values through the filters from blowing up. Sometime during the process is when we learned 32-bit integers were not good enough without a lot more intermediate overflow checks, expensive in terms of HW. This was before DSP chips were commonplace so the "DSP" used custom chips and bit-slice chips (AMD IIRC), but again I was only peripherally involved.
Of course, the digital guys wanted to know why we couldn't just design a 32-bit, 24 GS/s ADC and grab X-band signals directly. In ~1985. Pretty sure I couldn't do it even today.
Whatever, carry on - Don
It's maybe worth noting that DSP used in studios today is most likely using 64-bit float, which is considered adequate to avoid accumulating error / noise even after many, many repeated processing steps. 64-bit processing became common in DAWs years ago, shortly after 64-bit CPUs started to hit the consumer market.
It ends up being more involved than you'd think, for DSP. Sure, you need to manage the fixed point by shifting results, but it doesn't end there. If you multiply two numbers, you need twice as many bits for the result. You don't always need the least significant bits, but many times you do. Consider multiplying two small numbers inside a filter, to be later multiplied by a larger number. It's very easy for that first multiply to truncate to zero, then get multiplied by the larger number and produce a large error. So you typically need to implement that as having a larger accumulator that your sample size. The accumulator needs significant headroom too, because things like FIR filters can grow and shrink as they iterate to the result, and if bits fall off either end you're in trouble (especially on the large side, where the error will be huge). And, you need to saturate—you can't allow sums to roll over. For instance, the 24-bit fixed point 56K family has 24-bit data (binary point to left of msb), and a pair of 56-bit accumulators (48 bits to the right of the point, 8 bits to the left), and automatically saturates (clips) results when writing back to 24-bit memory.
But the 56K manages all that in hardware, everything executes in a cycle (and some parallelism too). For a general CPU, you'd have to code such a model, as well as dealing with a certain amount of awkwardness it beings.
Floating point operations don't need to saturate every operation, usually just at the final sample output. Also, while it truncates every multiply (results of multiplying two 25-bit mantissas has to fit back in 25-bit mantissa), and while that can be an issue with 32-bit float (it can kill you on intermediate operations is some filter configurations), 64-bit float is nearly as fast and has so many bits (54-bit mantissa), that truncation doesn't matter. And you don't need to manage a special accumulator with extra range, you can just multiply and add regular variables, saturate at the end.
danadam
Major Contributor
- Joined
- Jan 20, 2017
- Messages
- 1,014
- Likes
- 1,582
I hoped that would be 8-bit floats, which they weren't if I read that correctly. Oh well, let's be generous and assume they meant something else, so that at least one part of their sentence is true.
(Just to be clear to anyone reading, my facepalm was for they implying that 8-bits sounds flat and robotic)
danadam
Major Contributor
- Joined
- Jan 20, 2017
- Messages
- 1,014
- Likes
- 1,582
@BeerBear put the word "only" in quotes, so clearly it wasn't anything negative to him.It’s been tested to have 136dB dynamic range on that video. For some reason this was a negative
At the risk of joining @BeerBear in the "confused" group
, the quoted part there says about floating point type capabilities, not about any device performance.I'm not sure if anything can be inferred from it, but they specify maximum input level at +24 dBu (for TRS connection), which is 12.28 V. -136 dB is 1.9 µV, -192 dB is 3 nV. One of those two seems within the realms of possibility
.Yeah, it looks like capturing very low levels is hard. So even with multi-ADC designs we might not see dynamic ranges over 192dB, except in some lab gear maybe, IDK.I'm not sure if anything can be inferred from it, but they specify maximum input level at +24 dBu (for TRS connection), which is 12.28 V. -136 dB is 1.9 µV, -192 dB is 3 nV. One of those two seems within the realms of possibility
BTW, Sound Devices list the dynamic range specs for their MixPre series that record at 32bit float and it's 142dBA.
So over 144dB looks doable. And I guess 32bit FP is still going to be the default choice for the format, rather than 32bit integer. Not just because of visual/UI reasons that I mentioned before, but also because you can't losslessly process 32bit integer in a 32bit FP app (there are still such DAWs around, despite the trend towards 64bit FP).
- Joined
- Jul 23, 2019
- Messages
- 484
- Likes
- 512
I'm a trifle confused here. I understand that the sampled data is being converted to and stored as a 32-bit float. However, isn't the analog-to-digital converter only going to be working at 24-bit integer precision?
To put these things into perspective, here's someone's measurement results with these recorders. Notice how the RME using old technology showing the highest dynamic range and lowest distortion.
www.audiosciencereview.com
Another traditional recording interface:
www.audiosciencereview.com
Amir's measurements of Zoom F6:
www.audiosciencereview.com
This review of course, caused some troubles to Zoom's tech support department, but even after firmware update the results are still pathetic (the PDF links contain measurement results)
www.audiosciencereview.com
So while it is good to show off your high dynamic range figures when using unfair gain settings, all other aspects suffered. Even the illustration below gives some hints that the potential transition between two ADCs happens at typical operation levels and bad transition algorithms can induce additional artifacts, more bits, float or not, won't help.
RMAA Tests: (Welcome to add others!)
You need levels to remain the same for all devices to be able to compare. i did a bunch of these today, feeding from my rme adi2 pro fs and topping D70 at high (+24 dBU), mid (-2 dBU) and low levels (-14 dBU), to get a sense of preamp quality on those that have gain stages. a few were loopbacks...
www.audiosciencereview.com
Another traditional recording interface:
RMAA Tests: (Welcome to add others!)
Here's some RMAA tests. If anyone has other results to share as far as recording interface RMAA tests: you're welcome to post in this thread as more info on recording interfaces would be nice to know! -Or better yet send it in to be tested also :)
www.audiosciencereview.com
Amir's measurements of Zoom F6:
Zoom F6 Portable Field Recorder Review
This is a review and detailed measurements of the Zoom F6 battery operated, multi-channel portable balanced field recorder. It is on kind loan from a member and costs US $650 from Amazon including Prime shipping. I must say, the F6 is one sturdy feeling and looking recorder despite its small...
www.audiosciencereview.com
This review of course, caused some troubles to Zoom's tech support department, but even after firmware update the results are still pathetic (the PDF links contain measurement results)
Zoom F6 Portable Field Recorder Review
Although, in practice, it might be more fair to say that floating point has 144dB of dynamic range and 24 bit fixed has 138dB of dynamic range (23 bit equivalent), due mainly to the fact that the negative swinging side of the audio signal doesn't add to Dynamic Range. Haven't thought too much...
www.audiosciencereview.com
So while it is good to show off your high dynamic range figures when using unfair gain settings, all other aspects suffered. Even the illustration below gives some hints that the potential transition between two ADCs happens at typical operation levels and bad transition algorithms can induce additional artifacts, more bits, float or not, won't help.
Similar threads
