There are real implimentation problems when going below -90 -100dB. You start to see the ugly side of reality down there in the muck. Issues like windowing leakage, Unknown freqs that are not related to the fundamental, rfi/emi freqs that have affect which are not in the bins to see..... not to mention grounding paths for unwanted freqs during DUT measurements and system operation. On the high-side is more ADC non-linearity - some due to lowering of the operating voltage for portability apps... 3.3v operation; single-ended operation. Limited dynamic range that is linear.
Your second comment i high-lighted is always a fund. on my mind and a point I keep repeating with regard to models... and what one can or cannot percieve.
I had asked in another forum about whether anyone knew how to do parallel FFT. No one here seemed to think it was do-able.... no learning disorders there.... just linear thinking.
I read today the Chinese/Taiwan are doing it in another app for software radio. So, it can be done... wish we could get someone over here to apply it to FFT test and measurement.
(Design of Parallel FFT Based on FPGA in the Field of Software Radio. PIERS Proceedings, Taipei, March 25-28, 2013)
Thx-RNMarsh
Last edited:
that's an odd set of claims when anyone can pull up even free RMAA plots with -130 dB spot noise floor plots for better soundcards like LynxStudio, ESI Juli@ with -3 dB FS test sine...
MatLab, the free work-a-likes, Sage, Scientific Python all have extensively vetted FFT code, most can use 64 bit, some have higher accuracy options...
quite a bit of work is being done on FFT parallelism, GPU optimizations - quick search turns up things like: http://blogs.intel.com/intellabs/files/2012/11/SC12_FFT-pap147s4-final.pdf
and there is the question of audible significance - when do any of the "issues" you mention come within -20, -40?, even -60 dB of the repesentations used in >200kb psychoacoustic codecs - that are "audibly transparent" in very extensive blind testing?
MatLab, the free work-a-likes, Sage, Scientific Python all have extensively vetted FFT code, most can use 64 bit, some have higher accuracy options...
quite a bit of work is being done on FFT parallelism, GPU optimizations - quick search turns up things like: http://blogs.intel.com/intellabs/files/2012/11/SC12_FFT-pap147s4-final.pdf
and there is the question of audible significance - when do any of the "issues" you mention come within -20, -40?, even -60 dB of the repesentations used in >200kb psychoacoustic codecs - that are "audibly transparent" in very extensive blind testing?
Last edited:
You'd have to been following the other forum to get it. Not interested in number crunching with mega $ computing systems. Thank you.
Guess i just didnt ask the right group re parallel FFT's. 64 bit will be fine. When can I get something -parallel FFT/ADC hardware/software - to use on my PC?
And, where?
Thx-RNMarsh
Guess i just didnt ask the right group re parallel FFT's. 64 bit will be fine. When can I get something -parallel FFT/ADC hardware/software - to use on my PC?
And, where?
Thx-RNMarsh
Last edited:
Richard:
Terminology here is a trap. I think when you say parallel FFT you are really meaning parallel ADC, something quite different. Parallel FFT doesn't make a lot of sense unless different algorithms get different results (then one or the other is wrong). Or maybe you are looking at the spectrogram stuff that the SDR guys are playing with. The parallelism is to look at changes with time. It doesn't magically produce lower noise. What you are describing are limitations of real world hardware.
Below is a closer in view of an FFT I have shown before. Except for some hum leakage in the low frequencies and modulating the 1 KHz fundamental there is nothing above -160 dB (except the harmonics). The stuff below -160 dB is probably noise in the test setup. Probably not artifacts of FFT or even artifacts of the ADC. Getting noise or artifacts this low in a real audio system with connections to sources, speakers and power would be exceptionally hard. For what its worth the ADC plotted has a rated dynamic range of 123 dB.
Terminology here is a trap. I think when you say parallel FFT you are really meaning parallel ADC, something quite different. Parallel FFT doesn't make a lot of sense unless different algorithms get different results (then one or the other is wrong). Or maybe you are looking at the spectrogram stuff that the SDR guys are playing with. The parallelism is to look at changes with time. It doesn't magically produce lower noise. What you are describing are limitations of real world hardware.
Below is a closer in view of an FFT I have shown before. Except for some hum leakage in the low frequencies and modulating the 1 KHz fundamental there is nothing above -160 dB (except the harmonics). The stuff below -160 dB is probably noise in the test setup. Probably not artifacts of FFT or even artifacts of the ADC. Getting noise or artifacts this low in a real audio system with connections to sources, speakers and power would be exceptionally hard. For what its worth the ADC plotted has a rated dynamic range of 123 dB.
Attachments
Are you referring to TEF? Its not the same or even similar an FFT. TEF is a process for measuring a speaker ignoring stuff that didn't originate from the speaker. The basics have been translated into the digital domain and use FFT's quite well and get the same results much faster and easier.
The two situations are not comparable. Heaviside was not asking people to believe him on the basis of his 'guru' status or alleged track record, but on the basis that his maths and physics was correct. He was not saying "believe me"; he was saying "check what I say and see if it is true".john curl said:
Heaviside seemed to have poor social skills and often insulted people who disagreed with him. That caused both engineers and mathematicians to be inclined to dismiss his ideas without investigating them. As it happened, he was right and they were wrong. It would be foolish to conclude from this that arrogance is evidence of truth.
PS Parallel FFT appears to be merely a method for doing FFT. No new content so cannot change the results except due to floating point issues which are present in all numeric computation.
Last edited:
You asked the wrong people, write out the complete algorithm and insert multiply/accumulators and adders, then build the whole thing. It is now possible to actually put all that stuff in a small space and with realistic power and I/O.
Also please separate particular hardware implementations from the algorithm. No problem generating a digital sine wave and displaying the FFT to, let's see, double precision is about -300dB. In practice it's more in the ~ -200dB's.
I have posted the plots here several times. Go to the FFTW site at MIT just to see what is available and even FREE.
EDIT - The current Elektor series on Python is great. I have never found so much power in an open source package that has no catches in downloading and using, except using what is between your ears. There is the FFTW package available as a module. I highly recommend ActivePython 2.7 from Activestate, it even has a utility that goes out on the web to find modules that you might want to add, finds them AND any dependencies you might be lacking and installs everything with no hitches. Docs need a little seaching I guess geeks don't like to document their work.
Last edited:
This is another of your challenges, why would you want it? You get the same answer and current benchmarks on a good PC are quite amazing. Voltages are single valued at any given instant in time what would a parallel A/D do?
I hope that many reading this thread will take some time to read about Oliver Heaviside and his problems with Preece(the head of the government telegraph service).
This is a never-ending story of how someone notes something in nature, starts to think about it, theorizes and/or measures this phenomenon as best he can, and writes 'articles' about it in the journal that is most specialized in this subject area.
If you take a little time to read about Heaviside's history, you will find someone who is very interested in a subject, who reads up on just about everything about the subject, goes all out to understand it, yet when he tries to promote what he finds, he gets 'shut out' by the people in technical power at the time.
You will find that the situation is almost absurd, in that the guy seems to really know his math, has practical experience, and is virtually 'giving a new concept away' without even trying to patent it. Yet, he can be blocked from contributing to the field.
A good example of this was the theory of PIM, started (so far as I know) by Matti Otala in about 1978, at least, this is the first time that he discussed it with me.
I do know that he wrote a short, formal (full of math) paper for the AES in about 1980, hoping that it would appear in the AES Journal subsequently. They turned him down, yet he was a PhD professor in electronics engineering. Why? About the same reason that Heaviside was turned away from being published in 'The Electrician' after previously submitting a number of articles for the publication over several years before.
Those of you, with a little time, look up the debates between Preece and Heaviside, and judge for yourself. All that really happened is that 'progress' was put back a number of years. Happens all the time.
This is a never-ending story of how someone notes something in nature, starts to think about it, theorizes and/or measures this phenomenon as best he can, and writes 'articles' about it in the journal that is most specialized in this subject area.
If you take a little time to read about Heaviside's history, you will find someone who is very interested in a subject, who reads up on just about everything about the subject, goes all out to understand it, yet when he tries to promote what he finds, he gets 'shut out' by the people in technical power at the time.
You will find that the situation is almost absurd, in that the guy seems to really know his math, has practical experience, and is virtually 'giving a new concept away' without even trying to patent it. Yet, he can be blocked from contributing to the field.
A good example of this was the theory of PIM, started (so far as I know) by Matti Otala in about 1978, at least, this is the first time that he discussed it with me.
I do know that he wrote a short, formal (full of math) paper for the AES in about 1980, hoping that it would appear in the AES Journal subsequently. They turned him down, yet he was a PhD professor in electronics engineering. Why? About the same reason that Heaviside was turned away from being published in 'The Electrician' after previously submitting a number of articles for the publication over several years before.
Those of you, with a little time, look up the debates between Preece and Heaviside, and judge for yourself. All that really happened is that 'progress' was put back a number of years. Happens all the time.
I hope Matti is not as eccentric as Heavyside. Heavyside was interested in fixing a problem, looking at Ron's PIM measurements on tube amps I can only conclude that it is a good thing one of those euphonic colorations that don't exist..
". . . Those of you, with a little time, look up the debates between Preece and Heaviside, and judge for yourself. All that really happened is that 'progress' was put back a number of years. Happens all the time. "
Yes, yes, I fully agree.
I've been trying to push IC opamps as an ideal amplifying device for small signal audio here for quite a few years now and people just don't want to listen, and keep on sticking to those discrete circuits.
Happens all the time.
Yes, yes, I fully agree.
I've been trying to push IC opamps as an ideal amplifying device for small signal audio here for quite a few years now and people just don't want to listen, and keep on sticking to those discrete circuits.
Happens all the time.
Matti may have been dislexic, but he was no Heaviside. His assertion that high feedback caused PIM was completely wrong.
Once again, the does not mean that PIM does not exist, and that one of the mechanisms that can cause PIM is not feedback; it just means that large amounts of negative feedback and low open-loop bandwidth do not exacerbate PIM.
Cheers,
Bob
I might add there was no reason to be offended by Barrie Gilbert's comments. Take an ideal integrator and put a low order non-linearity at the input then close the loop. The equations can be written down by inspection, Barrie was simply blindsided by someone claiming the terms in the simple expansion of the solution was a "discovery".
We should double the size of the original thread by tomorrow, time to party?
We should double the size of the original thread by tomorrow, time to party?
Last edited:
Yes it is.
I usually get the best answeres or solutions from the best minds with an open-ended challenge without a lot of preconditions or pointing to my own direction. I want your/other's means and methods and creative ideas... I already have my limited thoughts. Yes, what advantage would parallel FFT offer? Computing speed? What else? Parallel computing of a different kind... multiple minds working on the same question.
Thx-RNMarsh
Last edited:
Well Scott, it seems to me that you have gotten to level 2. You know: It exists, but it's not important.
What Barrie's article, perhaps 15 years after Otala's AES preprint, did, was re-awaken interest in PIM.
What I would like people to understand is that people like Matti Otala are interested in WHY amplifiers sound the way they do, and Matti found that TIM was NOT the complete answer. New mechanisms had to be found to explain what was going on.
In 1978, Matti mentioned 'differential phase' as something to look into, and he wanted to pursue his series of AES papers in that direction.
You see, we, at that time, and even some of us, today, were dissatisfied with the sound quality of most amps and preamps at that time. Sometimes a really good tube amp seemed to fill the bill (as it sometimes does today), but solid state had a ways to go.
Both Matti and I (both together and independently) built very fast, wide open loop power amps in 1980. Matti did an amp for HK, and I did one for VMPS. Both were pretty good, but 'over the top' cost-wise, and neither was 'perfect' because we just did NOT know enough about materials (for instance) to make a 'perfect amp.
The IC of the day was the 5534, and I have to admit that it was a big improvement over previous IC's used for audio, BUT we still found that we could make discrete designs, better sounding.
We were always 'nagged' by: When we make low feedback, high speed amps, that they usually sounded 'better' than many stock amps with an IC in the front end and all kinds of negative feedback, making the MEASURED audio tests acceptable.
In all fairness, the PIM controversy fell away, until Barrie Gilbert's article. And I was amused that Barrie actually gave grudging acceptance to Matti Otala's work on TIM, as well. This opened up the controversy that continues today.
What Barrie's article, perhaps 15 years after Otala's AES preprint, did, was re-awaken interest in PIM.
What I would like people to understand is that people like Matti Otala are interested in WHY amplifiers sound the way they do, and Matti found that TIM was NOT the complete answer. New mechanisms had to be found to explain what was going on.
In 1978, Matti mentioned 'differential phase' as something to look into, and he wanted to pursue his series of AES papers in that direction.
You see, we, at that time, and even some of us, today, were dissatisfied with the sound quality of most amps and preamps at that time. Sometimes a really good tube amp seemed to fill the bill (as it sometimes does today), but solid state had a ways to go.
Both Matti and I (both together and independently) built very fast, wide open loop power amps in 1980. Matti did an amp for HK, and I did one for VMPS. Both were pretty good, but 'over the top' cost-wise, and neither was 'perfect' because we just did NOT know enough about materials (for instance) to make a 'perfect amp.
The IC of the day was the 5534, and I have to admit that it was a big improvement over previous IC's used for audio, BUT we still found that we could make discrete designs, better sounding.
We were always 'nagged' by: When we make low feedback, high speed amps, that they usually sounded 'better' than many stock amps with an IC in the front end and all kinds of negative feedback, making the MEASURED audio tests acceptable.
In all fairness, the PIM controversy fell away, until Barrie Gilbert's article. And I was amused that Barrie actually gave grudging acceptance to Matti Otala's work on TIM, as well. This opened up the controversy that continues today.
Yes it is.
Thx-RNMarsh
The original response you got still baffles me even Google turns up numerous examples. So you asked experts?
If you want to send TI a pile of money http://www.ti.com/lit/an/spra108/spra108.pdf
You misread my comment. Audiophile approved valve amps have loads of PIM so it must be important, for the next open design I propose we figure out how to add a tuned amount of low order PIM to an op-amp. This could be the holy grail in getting that tube sound out of solid state.
It seems parallel FFT per that example is a way to speed up the processing. While that is a useful thing, for off line analytics its not critical and no new information will be present at the output. It could make for a faster real time audio FFT but most current PC's are already capable of that, as are many tablets and phones.
Well, I guess it's somewhat like Doppler Distortion. Some kinds in some amount are OK, but that does not make FM distortion unimportant.
I hope that even more guidelines can be established on 'what works' vs what doesn't work as well. Kind of like automobiles, etc. I am not completely satisfied yet.
I hope that even more guidelines can be established on 'what works' vs what doesn't work as well. Kind of like automobiles, etc. I am not completely satisfied yet.
- Status
- Not open for further replies.