There was a pretty hot debate in the Blowtorch thread, regarding the importance of having a dual (matched) jfet pair at the input of Mr. Cordell's VinylTrak MM preamp. I maintained that dual jfets are here a useless expensive part, as any jfets from the same tube will do.
Mr. Cordell never showed any good arguments for using the LSK489 in the input stage. I was assuming that the core reason was to get some 2nd harmonic cancellation, therefore lowering the distortions. I did some simulations of the input stage, in isolation, that Mr. Cordell dismissed as "flawed", again without much explanations (I understand, his time is precious!), other than suggesting to simulate the entire preamp using his models.
A kind soul sent me the full VinylTrak schematic, so that I am now able to simulate the entire preamp. Unfortunately, a model for the LSK486 is not available on Mr. Cordell site, but only the (similar?) LS846. I had to assume that a LSK489 is a dual LS486 (which may or may not be true).
There is a flagrant error in the schematic, as published in LA: the U1B servo opamp has the inputs swapped. I think there might be another one, that seriously affects the PSRR, but we'll see about later.
By simulating the input stage in isolation, I have observed that the distortions quickly increase with the Idss mismatch, therefore an Idss1/Idss2=0.9 as guaranteed by the LSK489 data sheet is not good enough; a much tighter matching would be required. That conclusion is still valid; however it appears to hold only for the high transconductance BF862 that I used. Using the low transconductance LS846 model, the distortion variation, when changing the Beta ratio, is very small. This is in good concordance with Mr. Cordell's results, however it raises the same question: if distortions are barely depending on the input pair matching, then why bothering using an expensive matched pair at the input?
The answer is in simulating the entire preamp, and indeed Mr. Cordell is right: using a dual matched jfet input pair IS critical in his design (actually, more critical than we would like, a pair with matching at the data sheet limit is not good enough!), but for a completely different reason than distortions.
Unfortunately I can't put here the schematic - it's copyrighted and although I was told the article is available for downloading, I was unable to find it on the LA site. So I'll put here only the results, and quote the parts numbers in Mr. Cordell's original paper.
First, I parameterized the LS846 model by Beta (previously it was by Idss). This is not critical for the results, and it is more straightforward, as Beta is also a jfet model parameter (Idss is not). It's though a little more difficult to find the Beta values for which Idss is in the .09 ratio, per the LSK489 data sheet. Did some experiments, and found that by changing Beta by 0.05m, Idss ratio changes by 10%, so a Beta of +/-0.05m would be equivalent to a change of Idss ratio of +/-10%.
After fixing the servo error in the schematic (U1B has swapped inputs) the design settled nicely the bias. The distortions at 1KHz and 20mV input were indeed very low, about 0.03% and changes very little with the input pair mismatch. So the reason for using a matched input pair is not here.
According to Mr. Cordell's LS846 model, Beta is .9m (compare to the BF862 Beta of about 40m). I did a DC sweep with Beta=0.9+/-0.05m, corresponding to an Idss1/Idss2=1.1 ... 0.9 and I got the reason why the input pair matching is critical. The servo can correct only 80% of the DC range (so for an Idss1/Idss2=1.08 ... 0.7), which means the VinylTrak design relies on a better than data sheet matching of the input devices (which I am sure it happens in practice). Why is this happening? The reason is R20 (147k) which feeds the DC error signal. This resistor, together with the 75uS pole resistor (R18=3830 ohm) form a divider, which severely limits the U1B output correction range. The obvious solution would be to lower the R20=147k value, but is this possible? Not much, because Mr. Cordell certainly doesn't want the 75uS time constant, defined by R18, to be affected by the combination of R20 and the output impedance of the folded cascode. What I would do to fix this problem (and allow a much larger mismatch in the input stage pair, hence no need for the expensive dual part) would be to remove R18 and make R20=3830ohm. This change would imply that the U1B output impedance would stay much lower than 3830 ohm over the audio range, which may or not be true; regular opamp models (including the OPA134 used by Mr. Cordell) do not model the output impedance, so this has to be measured in the lab. Worst case, this would affect the RIAA precision, which is currently very good, I got better than 0.2%.
But one to another, Mr. Cordell was right and I was wrong. A dual matched pair is required in the VinylTrak MM input stage; but the core reason is the DC precision and the limited ability of the servo to correct. I think the design can be slightly changed (as described above) so that regular same tube devices could be used, but that's already another story.
Because I anyway captured the schematic, I tried to simulate the PSRR. The results with the values in the original schematic were not that great at low frequency. The core reasons are the low cap values in the current source/mirrors decoupling (C2=C4=0.1uF). Attached are the results for PSRR+ and PSRR- for (top to bottom) 0.1uF, 1uF, 10uF and 100uF. I would suggest a minimum of 10uF decoupling; this will lower the PSRR cutoff frequency at an acceptable value.
Mr. Cordell never showed any good arguments for using the LSK489 in the input stage. I was assuming that the core reason was to get some 2nd harmonic cancellation, therefore lowering the distortions. I did some simulations of the input stage, in isolation, that Mr. Cordell dismissed as "flawed", again without much explanations (I understand, his time is precious!), other than suggesting to simulate the entire preamp using his models.
A kind soul sent me the full VinylTrak schematic, so that I am now able to simulate the entire preamp. Unfortunately, a model for the LSK486 is not available on Mr. Cordell site, but only the (similar?) LS846. I had to assume that a LSK489 is a dual LS486 (which may or may not be true).
There is a flagrant error in the schematic, as published in LA: the U1B servo opamp has the inputs swapped. I think there might be another one, that seriously affects the PSRR, but we'll see about later.
By simulating the input stage in isolation, I have observed that the distortions quickly increase with the Idss mismatch, therefore an Idss1/Idss2=0.9 as guaranteed by the LSK489 data sheet is not good enough; a much tighter matching would be required. That conclusion is still valid; however it appears to hold only for the high transconductance BF862 that I used. Using the low transconductance LS846 model, the distortion variation, when changing the Beta ratio, is very small. This is in good concordance with Mr. Cordell's results, however it raises the same question: if distortions are barely depending on the input pair matching, then why bothering using an expensive matched pair at the input?
The answer is in simulating the entire preamp, and indeed Mr. Cordell is right: using a dual matched jfet input pair IS critical in his design (actually, more critical than we would like, a pair with matching at the data sheet limit is not good enough!), but for a completely different reason than distortions.
Unfortunately I can't put here the schematic - it's copyrighted and although I was told the article is available for downloading, I was unable to find it on the LA site. So I'll put here only the results, and quote the parts numbers in Mr. Cordell's original paper.
First, I parameterized the LS846 model by Beta (previously it was by Idss). This is not critical for the results, and it is more straightforward, as Beta is also a jfet model parameter (Idss is not). It's though a little more difficult to find the Beta values for which Idss is in the .09 ratio, per the LSK489 data sheet. Did some experiments, and found that by changing Beta by 0.05m, Idss ratio changes by 10%, so a Beta of +/-0.05m would be equivalent to a change of Idss ratio of +/-10%.
After fixing the servo error in the schematic (U1B has swapped inputs) the design settled nicely the bias. The distortions at 1KHz and 20mV input were indeed very low, about 0.03% and changes very little with the input pair mismatch. So the reason for using a matched input pair is not here.
According to Mr. Cordell's LS846 model, Beta is .9m (compare to the BF862 Beta of about 40m). I did a DC sweep with Beta=0.9+/-0.05m, corresponding to an Idss1/Idss2=1.1 ... 0.9 and I got the reason why the input pair matching is critical. The servo can correct only 80% of the DC range (so for an Idss1/Idss2=1.08 ... 0.7), which means the VinylTrak design relies on a better than data sheet matching of the input devices (which I am sure it happens in practice). Why is this happening? The reason is R20 (147k) which feeds the DC error signal. This resistor, together with the 75uS pole resistor (R18=3830 ohm) form a divider, which severely limits the U1B output correction range. The obvious solution would be to lower the R20=147k value, but is this possible? Not much, because Mr. Cordell certainly doesn't want the 75uS time constant, defined by R18, to be affected by the combination of R20 and the output impedance of the folded cascode. What I would do to fix this problem (and allow a much larger mismatch in the input stage pair, hence no need for the expensive dual part) would be to remove R18 and make R20=3830ohm. This change would imply that the U1B output impedance would stay much lower than 3830 ohm over the audio range, which may or not be true; regular opamp models (including the OPA134 used by Mr. Cordell) do not model the output impedance, so this has to be measured in the lab. Worst case, this would affect the RIAA precision, which is currently very good, I got better than 0.2%.
But one to another, Mr. Cordell was right and I was wrong. A dual matched pair is required in the VinylTrak MM input stage; but the core reason is the DC precision and the limited ability of the servo to correct. I think the design can be slightly changed (as described above) so that regular same tube devices could be used, but that's already another story.
Because I anyway captured the schematic, I tried to simulate the PSRR. The results with the values in the original schematic were not that great at low frequency. The core reasons are the low cap values in the current source/mirrors decoupling (C2=C4=0.1uF). Attached are the results for PSRR+ and PSRR- for (top to bottom) 0.1uF, 1uF, 10uF and 100uF. I would suggest a minimum of 10uF decoupling; this will lower the PSRR cutoff frequency at an acceptable value.