|
Post by jin167 on Jan 31, 2019 22:43:47 GMT -6
the link you gave me has little relevance to the topic that is being discussed here. Besides, as I said I'm aware of the difference between floating and fixed. Floating point has a better definition than a 'PROCESSING TECHNOLOGY'. Say that to any decent engineer or mathematician and see how they react to your remark. I'm not just talking about audio conversion, I'm talking about ADC in a broader sense. The paper below discusses the technical feasibility of floating point ADC. portal.research.lu.se/portal/files/4716761/1472266.pdfDon't think that your own definition of "conversion" applies to everyone. I'm an engineer and John is right. Floating point in layman's terms is merely a way to handle larger (or smaller depending on how you look at it) decimal values/numbers than a system could handle if they were working with static word length (bit width). If you do a math operation that results in a decimal number, but your system had a fixed bit width, you'd either have to round to an integer, or truncate your value. Either one loses precision or accuracy in the scheme of things. Your could also do more operations to figure out an integer result, but that wastes time and power. Yes, there's been talks of floating point specific converters, but they're nowhere to be found because they're impractical and ultimately unnecessary. Thank you, svart. I know you're an engineer and I highly value your opinion. I can agree with your definition/description of a floating point because it is exactly as how I understand it but wrapping it with a term 'PROCESSING TECHNOLOGY' is not really good enough for me. In this instance, John was adamant that a floating point converter is impossible by definition and I responded by saying that I do not have enough knowledge in dsp or converter design to conclude that a floating point converter is impossible but I have found number of papers regarding this topic suggesting their feasibility and as you have pointed out there have been some talks/research of floating point specific converters in the past but they never came to fruition for various reasons.
|
|
|
Post by johneppstein on Feb 1, 2019 7:52:25 GMT -6
While I, myself, am neither a coder nor a mathematician I CAN follow a paper and understand what it says. Here is such a paper from the well known company, the oddly named "Analog Devices" which makes various sorts of digital processing chips.
Hopefully you'l be able to understand the subject a bit better after reading it.
Again, floating point is a PROCESSING TECHNOLOGY, not a conversion technology. Analog to digital conversion, by nature produces a fixed point result.
I'm guessing that the papers you saw but didn't read involved conversion between fixed point and floating point within the processing environment and didn't have much of anything to do with analog to digital audio conversion. Back in the late 20th century and up to the early 2000s many CPUs did not actually have floating point built in and relied on exterrnal FPUs (Floating Point Processors). That may be what those papers were about.
Note that the word "conversion" means different things in different contexts.
the link you gave me has little relevance to the topic that is being discussed here. Besides, as I said I'm aware of the difference between floating and fixed. Floating point has a better definition than a 'PROCESSING TECHNOLOGY'. Say that to any decent engineer or mathematician and see how they react to your remark. I'm not just talking about audio conversion, I'm talking about ADC in a broader sense. The paper below discusses the technical feasibility of floating point ADC. portal.research.lu.se/portal/files/4716761/1472266.pdfDon't think that your own definition of "conversion" applies to everyone. You are joking, right?
It's not "my own definition of conversion". Conversion, as considered in this thread, is about the action performed by analog to digital audio converters and nothing else. The word "conversion" can mean a lot of different things in differrent con texts, from the process of changing one's religious allegience to turning a stock car into a hot rod. We are not discussing any of those other uses of the word here.
And the article I referred you to is EXACTLY what we're talking about.
|
|
|
Post by jin167 on Feb 1, 2019 9:02:44 GMT -6
the link you gave me has little relevance to the topic that is being discussed here. Besides, as I said I'm aware of the difference between floating and fixed. Floating point has a better definition than a 'PROCESSING TECHNOLOGY'. Say that to any decent engineer or mathematician and see how they react to your remark. I'm not just talking about audio conversion, I'm talking about ADC in a broader sense. The paper below discusses the technical feasibility of floating point ADC. portal.research.lu.se/portal/files/4716761/1472266.pdfDon't think that your own definition of "conversion" applies to everyone. You are joking, right?
It's not "my own definition of conversion". Conversion, as considered in this thread, is about the action performed by analog to digital audio converters and nothing else. The word "conversion" can mean a lot of different things in differrent con texts, from the process of changing one's religious allegience to turning a stock car into a hot rod. We are not discussing any of those other uses of the word here.
And tyhe article I referred you to is EXACTLY what we're talking about.
As long as you remain adamant in stating that ADC is ONLY possible in fixed point format, your definition of "conversion" should stick with you. And the link you have shared does not confirm your claim that conversion can ONLY happen in fixed point thus irrelevant to this topic. You have stated that FP ADC is impossible but there seem to be papers out there suggesting otherwise (BTW, I'm not interested in their practicality).
|
|
|
Post by johneppstein on Feb 1, 2019 13:20:10 GMT -6
I'm an engineer and John is right. Floating point in layman's terms is merely a way to handle larger (or smaller depending on how you look at it) decimal values/numbers than a system could handle if they were working with static word length (bit width). If you do a math operation that results in a decimal number, but your system had a fixed bit width, you'd either have to round to an integer, or truncate your value. Either one loses precision or accuracy in the scheme of things. Your could also do more operations to figure out an integer result, but that wastes time and power. Yes, there's been talks of floating point specific converters, but they're nowhere to be found because they're impractical and ultimately unnecessary. Thank you, svart. I know you're an engineer and I highly value your opinion. I can agree with your definition/description of a floating point because it is exactly as how I understand it but wrapping it with a term 'PROCESSING TECHNOLOGY' is not really good enough for me. In this instance, John was adamant that a floating point converter is impossible by definition and I responded by saying that I do not have enough knowledge in dsp or converter design to conclude that a floating point converter is impossible but I have found number of papers regarding this topic suggesting their feasibility and as you have pointed out there have been some talks/research of floating point specific converters in the past but they never came to fruition for various reasons. So far I'm only about 1/3 of my way through the paper you cited. It's pretty intense going. At this point it appears to me that it's simply describing techniques for designing an ADC that has an integrated FPU tacked on to the output. In such a case calling it a "floating point converter" is simply playing semantics - it's a fixed point converter that is run into an FPU that takes the remainder (which normally woulf be discarded) and converts it to floating point. At least that's how it appears to me at this point. It's going to take me a couple more days to get through the paper and I will probably have to read through it again and think about it for awhile.
Have you actually read the paper? Do you understand it?
My understanding comes from discussing this stuff with engineers and coders who actually work with this stuff. My impression is also that the stored data (written to disk or whatever) has to be in fixed point format. Perhaps Svart can comment on that?
|
|
|
Post by johneppstein on Feb 1, 2019 13:34:34 GMT -6
You are joking, right?
It's not "my own definition of conversion". Conversion, as considered in this thread, is about the action performed by analog to digital audio converters and nothing else. The word "conversion" can mean a lot of different things in differrent con texts, from the process of changing one's religious allegience to turning a stock car into a hot rod. We are not discussing any of those other uses of the word here.
And tyhe article I referred you to is EXACTLY what we're talking about.
As long as you remain adamant in stating that ADC is ONLY possible in fixed point format, your definition of "conversion" should stick with you. And the link you have shared does not confirm your claim that conversion can ONLY happen in fixed point thus irrelevant to this topic. You have stated that FP ADC is impossible but there seem to be papers out there suggesting otherwise (BTW, I'm not interested in their practicality). Well, if you're not interested in practicality this entire discussion is pointless. I'm talking about engineering, you're playing semantic games. And frankly, if that's all this is I have far more interesting science fiction books to read.
Am I wasting my time slogging through your paper?
|
|
|
Post by jin167 on Feb 1, 2019 20:09:51 GMT -6
As long as you remain adamant in stating that ADC is ONLY possible in fixed point format, your definition of "conversion" should stick with you. And the link you have shared does not confirm your claim that conversion can ONLY happen in fixed point thus irrelevant to this topic. You have stated that FP ADC is impossible but there seem to be papers out there suggesting otherwise (BTW, I'm not interested in their practicality). Well, if you're not interested in practicality this entire discussion is pointless. I'm talking about engineering, you're playing semantic games. And frankly, if that's all this is I have far more interesting science fiction books to read.
Am I wasting my time slogging through your paper?
I'm responding to your comment about FP ADC being IMPOSSIBLE to design/exist. There are papers out there suggesting its feasibility and I'm simply letting you know that they're there. Yes, I have read the paper and I'm reading it again because I have learned quiet a lot from it and discovering new things along the way. You're more than welcome to stop at any time it's not like I'm forcing you to read it. As I said, I'm not an expert in dsp or converter design but I can tell that you're not one neither so what makes you so certain that an FP ADC is impossible when you don't have sufficient knowledge to prove your claim? And what's up with dismissing someone's PhD thesis as science fiction? Do you have a PhD degree in electrical engineering? Again, I'm not interested in your own definition of FP ADC. If you want to create/define your own term, write a paper on it.
|
|
|
Post by christopher on Feb 1, 2019 22:14:00 GMT -6
32bit float format is not what this is doing. Sorry if I started a whole FP discussion, I should have clarified: what excites me slightly is the idea that maybe I could go back to not using FP and just referencing the actual captured wav, and playback that is exactly as-is out to the analog gear. Someday I'll hope to mix totally analog again.... (not even ITB faders) I'm planning 2035, lol, probably be some free SSL's by then... Its mostly the idea of being exactly the bits I tracked, not a rewritten improved version. So really 24bit is fine. Screw it... its higher res, I'll take 32.
|
|
|
Post by popmann on Feb 2, 2019 0:49:55 GMT -6
I think this all depends on how you define “define”.
|
|
|
Post by johneppstein on Feb 2, 2019 1:05:03 GMT -6
Well, if you're not interested in practicality this entire discussion is pointless. I'm talking about engineering, you're playing semantic games. And frankly, if that's all this is I have far more interesting science fiction books to read.
Am I wasting my time slogging through your paper?
I'm responding to your comment about FP ADC being IMPOSSIBLE to design/exist. There are papers out there suggesting its feasibility and I'm simply letting you know that they're there. Yes, I have read the paper and I'm reading it again because I have learned quiet a lot from it and discovering new things along the way. You're more than welcome to stop at any time it's not like I'm forcing you to read it. As I said, I'm not an expert in dsp or converter design but I can tell that you're not one neither so what makes you so certain that an FP ADC is impossible when you don't have sufficient knowledge to prove your claim? And what's up with dismissing someone's PhD thesis as science fiction? Do you have a PhD degree in electrical engineering? Again, I'm not interested in your own definition of FP ADC. If you want to create/define your own term, write a paper on it. Well, let's put it this way - Svart is not the only engineer who has told me that. Not even close.
I didn't say his thesis is SF - I said your interpretation of it is SF.
And let's get something straight right now - it's not my definition. what I'm saying isn't in any way radical, unconventional, or original. In fact, YOU are the first person I've ever encountered who has argued the point. It has been printed in manuals, textbooks, and articles explaining how floating point works in the DAW environment that are way too numerous to mention. Actually, for me it goes back to before DAWs for me, I believe I first encountered discussions of floating point back when I was hardware hacking on early personal computers and FPUs were separate chips back in the '80s in the old days of Silicon Valley.
Now, perhaps a "floating point converter" may be possible, but as I mentioned previously at the point I am in the paper it sure looks to me like the author is simply talking about adding an FPU to a standard fixed point converter and by the standards of those who design processors that isn't actually a floating point analog to digital converter. People who design CPUs, which are MUCH more intensive in their use of FP, still regard the onboard FPU in a modern CPU chip as a separate device from a design standpoint. The difference is that its on a section of the same die, not a discrete part. (And please don't start arguing semantics about the word "discrete".)
And even if a converter that is natively floating point might be possible there must be good reasons why such chips are not widespread. Like they're not practical, which is not much different from impossible from a practical engineering point of view.
|
|
|
Post by jin167 on Feb 2, 2019 1:38:56 GMT -6
I'm responding to your comment about FP ADC being IMPOSSIBLE to design/exist. There are papers out there suggesting its feasibility and I'm simply letting you know that they're there. Yes, I have read the paper and I'm reading it again because I have learned quiet a lot from it and discovering new things along the way. You're more than welcome to stop at any time it's not like I'm forcing you to read it. As I said, I'm not an expert in dsp or converter design but I can tell that you're not one neither so what makes you so certain that an FP ADC is impossible when you don't have sufficient knowledge to prove your claim? And what's up with dismissing someone's PhD thesis as science fiction? Do you have a PhD degree in electrical engineering? Again, I'm not interested in your own definition of FP ADC. If you want to create/define your own term, write a paper on it. Well, let's put it this way - Svart is not the only engineer who has told me that. Not even close.
I didn't say his thesis is SF - I said your interpretation of it is SF.
And let's get something straight right now - it's not my definition. what I'm saying isn't in any way radical, unconventional, or original. In fact, YOU are the first person I've ever encountered who has argued the point. It has been printed in manuals, textbooks, and articles explaining how floating point works in the DAW environment that are way too numerous to mention. Actually, for me it goes back to before DAWs for me, I believe I first encountered discussions of floating point back when I was hardware hacking on early personal computers and FPUs were separate chips back in the '80s in the old days of Silicon Valley.
Now, perhaps a "floating point converter" may be possible, but as I mentioned previously at the point I am in the paper it sure looks to me like the author is simply talking about adding an FPU to a standard fixed point converter and by the standards of those who design processors that isn't actually a floating point analog to digital converter. People who design CPUs, which are MUCH more intensive in their use of FP, still regard the onboard FPU in a modern CPU chip as a separate device from a design standpoint. The difference is that its on a section of the same die, not a discrete part. (And please don't start arguing semantics about the word "discrete".)
And even if a converter that is natively floating point might be possible there must be good reasons why such chips are not widespread. Like they're not practical, which is not much different from impossible from a practical engineering point of view.
I didn't say his thesis is SF - I said your interpretation of it is SF. Trying your best to dodge a bullet huh? Not even once have I expressed my interpretation of the paper on this forum and yet you claim that my interpretation of the paper is SF. Try harder. As I said, and for the last time, I'm questioning your claim that FP ADC is IMPOSSIBLE to design/exist. My quarrel is not about its practicality, I'm solely focusing on your claim that FP ADC is a design that is impossible to achieve. Just because something is impractical doesn't mean that it's impossible to design or build. It simply means there's no point in making one. And I was the first one to admit that I don't know enough about dsp or converter design to make any definitive statement but somehow you had enough confidence to state that an FP ADC is impossible to achieve despite not having sufficient technical knowledge (you can prove me wrong on this anytime. Perhaps a paper/article you have written on this subject?). This will be the last response from me. I can see that there is nothing to be gained from all this. Enjoy your day.
|
|
|
Post by BenjaminAshlin on Feb 2, 2019 4:05:59 GMT -6
Well that escalated quickly..
|
|
Deleted
Deleted Member
Posts: 0
|
Post by Deleted on Feb 2, 2019 7:12:33 GMT -6
Oh, guys, that was a very unproductive escalation. A floating point converter that spits out 32 or 64 or even 1024 bits is still not floating point in a mathematical sense, it is just different precisions of digitally processable representations of floating point numbers ... If we start to play around with these definitions comparing apples to oranges randomly, a floating point converter is pointless, because, if it is floating point in a mathematical sense, there is no digitalization at all. (Floating point numbers have no steps). These discussions about FP-ADCs have a mere academical value but are irrelevant to audio engineering. Svart already pointed to it and he is perfectly right. And yes, I have an academical and engineering training and worked also as a coder... Please, throwing in papers that have absolutely no practical relation to audio engineering just to prove there are different definitions to words does not make sense at all IMHO.
|
|
|
Post by Guitar on Feb 2, 2019 7:21:44 GMT -6
A classic online pissing match. It happens to all of us from time to time.
|
|
|
Post by johneppstein on Feb 2, 2019 18:09:58 GMT -6
Well that escalated quickly.. Not my fault - or problem.
If he's going to stop bugging me about it that's great. I do kinda why he hasn't been bugging Svart about if after Svart told him I was right, but frankly I think the whole discussion is a waste of time.
|
|
|
Post by johneppstein on Feb 2, 2019 18:11:09 GMT -6
Oh, guys, that was a very unproductive escalation. A floating point converter that spits out 32 or 64 or even 1024 bits is still not floating point in a mathematical sense, it is just different precisions of digitally processable representations of floating point numbers ... If we start to play around with these definitions comparing apples to oranges randomly, a floating point converter is pointless, because, if it is floating point in a mathematical sense, there is no digitalization at all. (Floating point numbers have no steps). These discussions about FP-ADCs have a mere academical value but are irrelevant to audio engineering. Svart already pointed to it and he is perfectly right. And yes, I have an academical and engineering training and worked also as a coder... Please, throwing in papers that have absolutely no practical relation to audio engineering just to prove there are different definitions to words does not make sense at all IMHO. Thank you.
|
|
|
Post by johneppstein on Feb 2, 2019 18:12:50 GMT -6
This will be the last response from me. I can see that there is nothing to be gained from all this. Enjoy your day. Good. Thank you. Have a good day.
|
|
|
Post by Omicron9 on Feb 5, 2019 11:20:05 GMT -6
Politics. Religion. ADC.
|
|
mhep
Full Member
Posts: 36
|
Post by mhep on Feb 6, 2019 18:29:09 GMT -6
Tell me a source you need to record with >144dB dynamic range.
384kS/s is equally without audible merit. What source will you need to record that reaches up to 192kHz. If you don't believe Nyquist, tell me what company needs to make filtering so badly they need an extra 300kHz to work with.
We've moved past smoke and mirrors here.
|
|
|
Post by svart on Feb 7, 2019 12:40:42 GMT -6
Thank you, svart. I know you're an engineer and I highly value your opinion. I can agree with your definition/description of a floating point because it is exactly as how I understand it but wrapping it with a term 'PROCESSING TECHNOLOGY' is not really good enough for me. In this instance, John was adamant that a floating point converter is impossible by definition and I responded by saying that I do not have enough knowledge in dsp or converter design to conclude that a floating point converter is impossible but I have found number of papers regarding this topic suggesting their feasibility and as you have pointed out there have been some talks/research of floating point specific converters in the past but they never came to fruition for various reasons. So far I'm only about 1/3 of my way through the paper you cited. It's pretty intense going. At this point it appears to me that it's simply describing techniques for designing an ADC that has an integrated FPU tacked on to the output. In such a case calling it a "floating point converter" is simply playing semantics - it's a fixed point converter that is run into an FPU that takes the remainder (which normally woulf be discarded) and converts it to floating point. At least that's how it appears to me at this point. It's going to take me a couple more days to get through the paper and I will probably have to read through it again and think about it for awhile.
Have you actually read the paper? Do you understand it?
My understanding comes from discussing this stuff with engineers and coders who actually work with this stuff. My impression is also that the stored data (written to disk or whatever) has to be in fixed point format. Perhaps Svart can comment on that?
Strange, I never got a notice that you had tagged me on this.. Anyway, I have not read the paper you guys are discussing, but it's my understanding that FP-ADCs suffer from a specific problem relating to multiple signals which have large amplitude deltas(differences). Since the basis of a FP-ADC is (in layman's terms) a sliding window of lower resolution that reacts to the highest dynamic range of signals to effectively encode a higher dynamic range, if you had two or more spectrally uncoupled signals (in frequency domain this would look like two sines at different frequencies), the FP-ADC would react to the largest magnitude signal and the smaller of the two would have degraded acquisition accuracy. That also doesn't take into account the decoupling of SNR and DR and the varying quantization noise over signal magnitude. The type of dither needed to help squash the varying quantization noise would increase noise floors higher than we're used to seeing with what we expect from modern dither technique. It's just more practical to have SNR and DR coupled and use a higher bit-depth device with a known datarate. While we waste some bits trying to get to higher resolutions and the data is excessive, it's easier to deal with these drawbacks than those of a true FP-ADC. There's been papers on logarithmic ADC's and true FP-ADCs for decades now, but i don't think anyone has been able to solve all the riddles yet. I'm sure that if they do, we'll see a healthy adoption from the industry much like we did with delta-sigma devices.
|
|
|
Post by shoe on Feb 7, 2019 18:54:15 GMT -6
Tell me a source you need to record with >144dB dynamic range. 384kS/s is equally without audible merit. What source will you need to record that reaches up to 192kHz. If you don't believe Nyquist, tell me what company needs to make filtering so badly they need an extra 300kHz to work with. We've moved past smoke and mirrors here. Maybe you need 32 bits to compensate for all the headroom those sweet 192khz tones are soaking up.
|
|
|
Post by Bob Olhsson on Feb 9, 2019 10:46:52 GMT -6
There is a simple test that reveals all. Make a recording PEAKING at -20 with a 24 bit converter. Then play it back with some processing to turn the volume up. What I found, quite by accident, is that it will usually sound surprisingly better with no more apparent noise.
|
|