On 20 Dec 2003 15:39:58 -0800, santosh.nath@ntlworld.com (santosh nath) wrote:>Though it is not very good (rather stupid) to ask the following >questions; I thought this could be a pre-Xmas fun! > >1. What is the largest FFT size(point) ever used so far and >where(Application)?Largest I've used is 2^19 point, for testing a pink noise generator. I did 1024 of them and averaged the results. I think it took almost an hour to run on a Sparc. http://www.firstpr.com.au/dsp/pink-noise/allan-2/spectrum2.html>2. What is the largest QAM constellation ever used commercially so far >and where(Application)? ?I once worked for a company that had a 1024 QAM modem (for undersea cables). It was developed in about 1992.>3. What is the largest FIR filter taps used so far and >where(Application)?.I once made a 240 pole *analog* low pass filter. Regards, Allan.
The system that requires largest!
Started by ●December 20, 2003
Reply by ●December 21, 20032003-12-21
Reply by ●December 22, 20032003-12-22
Richard Owlett <rowlett@atlascomm.net> wrote in message news:<vubvmrbkmbmvb6@corp.supernews.com>...> Rick Lyons wrote: > >>Also, later I suspect I will want to compare spectra of data with > >>different window widths while having the same resolution in the > >>frequency domain. Each spectrum will be "normalized against itself", > >>either by saying the largest peak or the value at a certain frequency > >>is 0 dB. How? > > > > > > Hi Richard, > > Humm, I didn't understand your last question. > > Different window widths (i.e., the length of > > your time sequences) will result in different > > freq-domain resolutions. That's assuming the sample > > rate is the same for the various time-domain sequences. > > > > Obvious now that you mention it ;{ > How do you calculate the nominal center of each point in the spectrum? > [ I did a dimensional analysis of the formula I used and came up with > trash ;[What do you mean by "nominal center"? One way of getting comparable spectra is to use different window lengths and zero-pad the shorter (windowed) data frames to the longest frame length. In that way each spectrum will have the same "apparent resolution" and can be plotted easily in one plot and also be comprehensable.> > You can plot spectra where the largest spec magnitude > > sample is always at 0 dB by: > > [snip] > > At least I had that much right.Rick is right, but be aware that if the peak of the spectrum varies rapidly between frames, you will find an oscillating spectrum. If you intend to plot some variation of the spectrogram, things can become very confusing. It might be better to use a fixed reference in computing the dB scale. The simplest way would be to compute the dB scale as Rick suggested, but where you skip points 3 and 4, which means you set the dB reference to 1. Rune
Reply by ●December 22, 20032003-12-22
Allan Herriman wrote: ...> I once made a 240 pole *analog* low pass filter. > > Regards, > Allan.That's amazing! (I assume it worked.) Jerry -- Engineering is the art of making what you want from things you can get. �����������������������������������������������������������������������
Reply by ●December 22, 20032003-12-22
On Sun, 21 Dec 2003 14:28:49 -0500, Jerry Avins <jya@ieee.org> wrote:> Jerry Avins wrote: > >> >> >> What's with the preoccupation with biggest and mostest? Financial >> analysts perform FFTs and transversal FIRs using tens of thousands of >> bins and taps. I haven't kept track of the actual numbers. > > Correction: hundreds of thousands. > > Jerry > -- > Engineering is the art of making what you want from things you can get. > ����������������������������������������������������������������������� >Jerry, you have links for that? I ran a 4MB fft in LabView once. Started it, went to lunch, went to a long design review . . . Got back just in time (three hours on a 200MHz Pentium)
Reply by ●December 22, 20032003-12-22
On Sun, 21 Dec 2003 14:14:58 -0600, Richard Owlett <rowlett@atlascomm.net> wrote: (snipped)> >Obvious now that you mention it ;{ >How do you calculate the nominal center of each point in the spectrum? >[ I did a dimensional analysis of the formula I used and came up with >trash ;[ >Oops, I'm not sure what you mean by "center of a point". [-Rick-]
Reply by ●December 22, 20032003-12-22
Charles Krug wrote:> On Sun, 21 Dec 2003 14:28:49 -0500, Jerry Avins <jya@ieee.org> wrote: > >>Jerry Avins wrote: >> >> >>> >>>What's with the preoccupation with biggest and mostest? Financial >>>analysts perform FFTs and transversal FIRs using tens of thousands of >>>bins and taps. I haven't kept track of the actual numbers. >> >>Correction: hundreds of thousands. >> >>Jerry >>-- >>Engineering is the art of making what you want from things you can get. >>����������������������������������������������������������������������� >> > > > Jerry, you have links for that? > > I ran a 4MB fft in LabView once. Started it, went to lunch, went to a > long design review . . . > > Got back just in time (three hours on a 200MHz Pentium)I'm sorry, no. There was a financial analyst who posted here for a while, but his name escapes me, an interesting if contentious fellow who stimulated discussion by casting different lights on things (especially stability). Although he was necessarily closed-mouthed about his techniques, he mentioned overnight FFTs on a fast mainframe. Jerry -- Engineering is the art of making what you want from things you can get. �����������������������������������������������������������������������
Reply by ●December 22, 20032003-12-22
"santosh nath" <santosh.nath@ntlworld.com> wrote in message news:6afd943a.0312201539.555ad3a6@posting.google.com...> Though it is not very good (rather stupid) to ask the following > questions; I thought this could be a pre-Xmas fun! > > 1. What is the largest FFT size(point) ever used so far and > where(Application)? > 2. What is the largest QAM constellation ever used commercially so far > and where(Application)? ? > 3. What is the largest FIR filter taps used so far and > where(Application)?.It's important to realize that (1) and (3) could be connected because of fast convolution using fft/multiply/ifft. Long FIR = Longer FFT. The larger the arrays, the more likely this approach will be used! I once worked with physicists who were doing acoustic modeling of large, complex bodies for sonar. They would convolve the impulse response generated by their model with the sonar transmit pulse in order to see what echoes from this object would look like. The spacing of the individual scatterers wasn't regular to say the least! And, the spacing from one scatterer to the next was many, many wavelengths. Generating the echoes using time domain convolution was way too expensive in compute time. So, I told them they might try this: 1) Discretize the location of the scatterers in time according to some criterion that would place them "close enough" to the location they'd decided to use. Maybe 1/20th of a wavelength and maybe greater. The former makes more points - maybe an increase in the number of points by 200. 2) Zero fill all the other spatial sample points. So, starting with a few hundred scatterers, this could result in 40,000 points or so. 3) Use fast convolution - more or less doubled the number of points in this case - so the FIR filter could have had a length of 40,000 and the FFT a length of 80,000 to 100,000. Oh yes, the speed was improved by a factor of 10! Now that I think about it, I'm not sure why because the number of computed points increase had to be around a factor of 200 or so. Maybe we got rid of another inefficiency in the process! Fred
Reply by ●December 22, 20032003-12-22
"santosh nath" <santosh.nath@ntlworld.com> wrote in message news:6afd943a.0312201539.555ad3a6@posting.google.com...> Though it is not very good (rather stupid) to ask the following > questions; I thought this could be a pre-Xmas fun! > > 1. What is the largest FFT size(point) ever used so far and > where(Application)? > 2. What is the largest QAM constellation ever used commercially so far > and where(Application)? ? > 3. What is the largest FIR filter taps used so far and > where(Application)?. > > Real time constraint should be applied if needed.Like you already pointed, it doesn't seem like the answers would be of much use other than for some boasting perhaps. We use more than a million point FFT - not sure exactly how many but certainly < 10 million. Application is spectral analysis. We also use pretty long FIR filters - longest I have running right now is a little under 1000 taps. Largest commercially used QAM i've heard of is 256 QAM in cable or DVB applications. Cheers Bhaskar
Reply by ●December 22, 20032003-12-22
I lurk on comp.lang.research . A frequent topic of discussion is detecting "silence" [ i.e. pauses between words/phrases/sentences ] Being one who talks a lot ;] I have a "gut" feel that *each* of those might be of fairly consistent duration for a particular speaker. Ignoring for the moment the current push for "real time"/"instant" recognition. Given a speech sample of tens of seconds, should/could/would an FFT of the *ENTIRE* utterance give indications of where and how long pauses were?
Reply by ●December 22, 20032003-12-22
"Jerry Avins" <jya@ieee.org> wrote in message news:3fe71c4c$0$4736$61fed72c@news.rcn.com...>> I'm sorry, no. There was a financial analyst who posted here for a > while, but his name escapes me, an interesting if contentious fellow who > stimulated discussion by casting different lights on things (especially > stability). Although he was necessarily closed-mouthed about his > techniques, he mentioned overnight FFTs on a fast mainframe. >Jerry, et. al., IIRC it was Andrew. I thought it was more than coincidence that his posting here stopped when the stock market tanked. Clay
