BTW I just checked that the FQP5N50 (post #7) is already obsolete.
Only FQP5N60C is still active.
Would the latter still be a good complementary to the FQP3P50 ?
The datasheet curves are on log scale and really difficult to read.
Without a device to measure, it is not so easy to compare.
Patrick
Only FQP5N60C is still active.
Would the latter still be a good complementary to the FQP3P50 ?
The datasheet curves are on log scale and really difficult to read.
Without a device to measure, it is not so easy to compare.
Patrick
My apologies, I mixed up the part numbers. I was talking about the FDP5N50NZ and the FQP3P50. Looking at the datasheet, they roughly have similar transconductance (albeit at different drain currents), input capacitance, and threshold voltage. The RDSON is very different, but that only matters for switching applications, not here.
The MOSFETs in this design have rather large source resistors, so the degeneration is strong, and the difference in transconductance does not matter that much.
Here is the P-channel: https://www.fairchildsemi.com/datasheets/FQ/FQP3P50.pdf
Here is the N-channel : https://www.fairchildsemi.com/datasheets/FD/FDP5N50NZ.pdf
Especially the P-channel is a device that I would consider a sleeper, as it is an older design and - in comparison - much more "linear" than modern (NMOS) transistors, and in contrast to tube amplifiers, where a p-channel tube does not exist, this gives quite a degree of freedom. The NMOS on the other hand is a much more recent planar device, still much more benign than more advanced superjunction or trench devices, which are really only good for switching applications, and can have issues in linear mode, up until destruction - handle with care....
What counts more is the input stage, as Patrick rightly mentioned, and that is where I would be looking for further improvement.
The MOSFETs in this design have rather large source resistors, so the degeneration is strong, and the difference in transconductance does not matter that much.
Here is the P-channel: https://www.fairchildsemi.com/datasheets/FQ/FQP3P50.pdf
Here is the N-channel : https://www.fairchildsemi.com/datasheets/FD/FDP5N50NZ.pdf
Especially the P-channel is a device that I would consider a sleeper, as it is an older design and - in comparison - much more "linear" than modern (NMOS) transistors, and in contrast to tube amplifiers, where a p-channel tube does not exist, this gives quite a degree of freedom. The NMOS on the other hand is a much more recent planar device, still much more benign than more advanced superjunction or trench devices, which are really only good for switching applications, and can have issues in linear mode, up until destruction - handle with care....
What counts more is the input stage, as Patrick rightly mentioned, and that is where I would be looking for further improvement.
I spent the last couple of days building a model exactly to that in post #7, in order to understand your design choices better.
The models of the power FETs were from Fairchild, the BJTs from Cordell.
There were no models from Fairchild for the JFETs, but the datasheet hinted that PF5102 is identical to J111.
So I took some old models of J111 and J270 from my library instead (Version 00).
Then I tried to play around with different parameters and values.
But it seems that you have got everything pretty optimum.
So many congratulations.
I substituted the JFETs to 2SK170/2SJ74s (Version 0). And as you can see K3 & K5 are somewhat lower.
(For direct comparison I set the output to +/-70V per single ended output.)
Bandwidth is much lower though at 200kHz instead of over 1MHz.
This is actually not a bad thing.
At 1MHz, a bit of parasitic inductance in the layout will set things into oscillation.
One can also now use a 1k output resistor with almost no loss in bandwidth.
I expect also the Toshibas to be quite a bit lower noise than the J-variants, especially the P device.
One thing I found interesting is that the 100R resistor across the input differential pair has a dominant effect on odd harmonics.
It reduces K3 & K5 by almost a factor of 3 at the same output level, but at the expense of closed loop gain.
So you will need >1Vrms at the input to get the amp to clip.
The other thing is that using a common 1mA potential divider for the BJT cascode base has no detrimental effect compared to separate ones.
All simulations of course.
However, in case of marginal stability, I would perhaps rather use a Zobel than bandwidth limiting in the feedback loop.
(Just personal choice, no scientific arguments.)
Hope you'll be able to report some listening impressions soon.
(And it seems that one cannot get KSC3503-E anywhere for some reasons.)
Patrick
.
The models of the power FETs were from Fairchild, the BJTs from Cordell.
There were no models from Fairchild for the JFETs, but the datasheet hinted that PF5102 is identical to J111.
So I took some old models of J111 and J270 from my library instead (Version 00).
Then I tried to play around with different parameters and values.
But it seems that you have got everything pretty optimum.
So many congratulations.
I substituted the JFETs to 2SK170/2SJ74s (Version 0). And as you can see K3 & K5 are somewhat lower.
(For direct comparison I set the output to +/-70V per single ended output.)
Bandwidth is much lower though at 200kHz instead of over 1MHz.
This is actually not a bad thing.
At 1MHz, a bit of parasitic inductance in the layout will set things into oscillation.
One can also now use a 1k output resistor with almost no loss in bandwidth.
I expect also the Toshibas to be quite a bit lower noise than the J-variants, especially the P device.
One thing I found interesting is that the 100R resistor across the input differential pair has a dominant effect on odd harmonics.
It reduces K3 & K5 by almost a factor of 3 at the same output level, but at the expense of closed loop gain.
So you will need >1Vrms at the input to get the amp to clip.
The other thing is that using a common 1mA potential divider for the BJT cascode base has no detrimental effect compared to separate ones.
All simulations of course.
However, in case of marginal stability, I would perhaps rather use a Zobel than bandwidth limiting in the feedback loop.
(Just personal choice, no scientific arguments.)
Hope you'll be able to report some listening impressions soon.
(And it seems that one cannot get KSC3503-E anywhere for some reasons.)
Patrick
.
Checked the datasheet, and the FQP5N60C looks closer from a capacitance point of view. The transconductance is even higher, though a) it is hard to say what the value is at 10mA as compared to 1A, b) with the source degeneration it doesnt matter. Thanks for the tip!!
I am using the 3503/1381 from Fairchild, since I have a bunch of them
I spent the last couple of days building a model exactly to that in post #7, in order to understand your design choices better.
The models of the power FETs were from Fairchild, the BJTs from Cordell.
There were no models from Fairchild for the JFETs, but the datasheet hinted that PF5102 is identical to J111.
So I took some old models of J111 and J270 from my library instead (Version 00).
Then I tried to play around with different parameters and values.
But it seems that you have got everything pretty optimum.
So many congratulations.
I feel honoured! (I really do). Thank you for the flowers!
ltwiki.org has long files with all kinds of bipolar and JFET models, I simply used their files and - voila - the PF5102 is included there, along with many others. Have not checked how good the models are, they seemed pretty close to me, however that says nothing once you do distortion or noise simulations.
That is interesting, I thought the bandwith limitation is mostly due to the gm of the input device and gate capacitance of the power device, and as the K170 is stronger the bandwidth should be even higher... have to check that datasheet again, and will try and rerun your simulations.
On the output resistor, at worst case 250V max output voltage there will be a lot of power dissipation in that resistor, and the amplifier will leave class A . So a Zobel might be a good idea.
Yes, I designed it to be driven by a line stage, so a input voltage of several volts is available. In fact, my "old" line stage has way too much gain, and I have to run it with the volume control always way down. This one will go inside my new line stage, where the gain will be lower but a few volts will be available as well. Is this a design with a lot of feedback? Perhaps yes, more than the original F5, but designed to a purpose, and it is mostly local feedback (source degeneration in the input and output stages), whereas the global feedback is rather weak.
Yes, that is coming soon. A first rough check did sound nice, with lots of details, but this was just a first check, with wild cabling and no warm-up.
Spice says that adding a base resistor to the BJT Cascodes improves stability.
3k according to my models.
Might not be real, but worth a try if not too much fiddling required.
Current is about 25µA. A 0805 SMD will do.
I think you are using so much degeneration in the second stage that the MOSFETs are no longer determining the bandwidth.
The front end appears to be.
The ratio of degen R & 1/Yfs determines how much of the capacitances is effectively seen at the gate.
Patrick
3k according to my models.
Might not be real, but worth a try if not too much fiddling required.
Current is about 25µA. A 0805 SMD will do.
I think you are using so much degeneration in the second stage that the MOSFETs are no longer determining the bandwidth.
The front end appears to be.
The ratio of degen R & 1/Yfs determines how much of the capacitances is effectively seen at the gate.
Patrick
I know from previous experience that cascodes play a significant part in frequency response.
I also wanted to check how much current was going through the gates of the cascode.
As I was worried that the low hfe of high voltage BJTs will bring additional distortion.
So I added some large base resistors like 10k just to see what happens.
With 3k, the response of your original circuit almost looks like those using Toshiba JFETs but without base resistors.
You can use my models in post #24 and add some R's just to try out.
Patrick
I also wanted to check how much current was going through the gates of the cascode.
As I was worried that the low hfe of high voltage BJTs will bring additional distortion.
So I added some large base resistors like 10k just to see what happens.
With 3k, the response of your original circuit almost looks like those using Toshiba JFETs but without base resistors.
You can use my models in post #24 and add some R's just to try out.
Patrick
Tried your simulation files, many thanks for all the work you put in there!! I am not sure about the output loading though, I was under the impression that the two outputs will drive the two stators, so your loading seems right (150pF between the outputs), tried it with 150pF to ground and it did not change matters much... So much for thinking about stuff while I am typing
There is huge debates ongoing regarding the upper bandwidth. I remember a statement from John Curl, where he argued that a decent amplifier should have 100V/us or so to "sound good". Taken out of context , ok but that was my design target. And according to my original simulations thats where I settled. It would still be interesting to understand where that lower upper corner frequency is coming from, with the K170 having 3x the transconductance of the PF5102.....
There is huge debates ongoing regarding the upper bandwidth. I remember a statement from John Curl, where he argued that a decent amplifier should have 100V/us or so to "sound good". Taken out of context , ok but that was my design target. And according to my original simulations thats where I settled. It would still be interesting to understand where that lower upper corner frequency is coming from, with the K170 having 3x the transconductance of the PF5102.....
150p between outputs equals 2x 300p to Gnd.
And that is a heavier load than 150p to Gnd.
The 100V/µs is not a holy value, as it is amplitude dependent.
It has totally different meaning whether it is line level, power amp or electrostatic.
Bandwidth is on the other hand not amplitude dependent.
Of course it does no harm to have 1MHz bandwidth, as long as you have good stability.
But at such high frequencies, layout becomes important and parasitic track inductances can make things unstable.
And also not always easy to track.
Using bandwidth limiting is easy way out.
But if the problem is too much HF gain, I rather kill that by loading it with a Zobel.
And you can argue which approach is the better one ......
Patrick
And that is a heavier load than 150p to Gnd.
The 100V/µs is not a holy value, as it is amplitude dependent.
It has totally different meaning whether it is line level, power amp or electrostatic.
Bandwidth is on the other hand not amplitude dependent.
Of course it does no harm to have 1MHz bandwidth, as long as you have good stability.
But at such high frequencies, layout becomes important and parasitic track inductances can make things unstable.
And also not always easy to track.
Using bandwidth limiting is easy way out.
But if the problem is too much HF gain, I rather kill that by loading it with a Zobel.
And you can argue which approach is the better one ......
Patrick
Thinking about it, it may be better to limit the bandwidth through the cascodes, rather than with capacitors across the 24k feedback resistors (which I did), simply because this solves the problem where it is created, rather than putting in an overall cure - just a stomach feeling.
Yes, layout is a important thing, as I found out the hard way doing the Paradise PCBs. Nowadays I put ground planes and proper decoupling where needed, but not on these amplifiers. They have a rather large ground connection down the middle of the PCB, but it seems to be not fully effective. Maybe one day I'll redo the layout...
I have now disassembled the breadboard to put it into the case, but I do have a third "test" amplifier, where I can try some modifications. Problem is, these ideas and inputs are really good advice, so I may end up disassembling the PCBs from the heatsinks, to modify them - lots of work, sigh...
Yes, layout is a important thing, as I found out the hard way doing the Paradise PCBs. Nowadays I put ground planes and proper decoupling where needed, but not on these amplifiers. They have a rather large ground connection down the middle of the PCB, but it seems to be not fully effective. Maybe one day I'll redo the layout...
I have now disassembled the breadboard to put it into the case, but I do have a third "test" amplifier, where I can try some modifications. Problem is, these ideas and inputs are really good advice, so I may end up disassembling the PCBs from the heatsinks, to modify them - lots of work, sigh...
The F5X / F5X Pre PCB has no Ground plane, just a small central Star Gnd point in the middle.
It works close to 1MHz with stable square wave response.
I do not believe that a flooding Gnd plane is universally good.
You may not always want parasitic capacitances between signal paths and Gnd.
Treat them as prototypes.
You will want to rebuild anyhow after all these fine optimisations.
There is no fun if it is always first time right.
Patrick
It works close to 1MHz with stable square wave response.
I do not believe that a flooding Gnd plane is universally good.
You may not always want parasitic capacitances between signal paths and Gnd.
Treat them as prototypes.
You will want to rebuild anyhow after all these fine optimisations.
There is no fun if it is always first time right.
Patrick
There were a few things that intrigued me, and after a couple hours of simming around I finally saw what it was.
In your sims, the current in the output devices is a lot higher, 150mA or so, whereas in mine the current is 17mA. This is due to the servos which I did not publish (yet), see attached. These keep both halves at pretty much the same current give or take, and the feedback loop will then do the little rest.
I was wondering where the large difference between upper and lower half was coming from,, and in a "unregulated" setup that is clear. Funny enough, distortion is up at .06% but base resistors on the bipolars have no more effect now.... odd....
In your sims, the current in the output devices is a lot higher, 150mA or so, whereas in mine the current is 17mA. This is due to the servos which I did not publish (yet), see attached. These keep both halves at pretty much the same current give or take, and the feedback loop will then do the little rest.
I was wondering where the large difference between upper and lower half was coming from,, and in a "unregulated" setup that is clear. Funny enough, distortion is up at .06% but base resistors on the bipolars have no more effect now.... odd....
The V00 that I posted was a carbon copy of your schematics.
The second stage bias is roughly determined by the drain resistor of the JFET and the source resistor of the MOSFET.
i.e. about a factor of 10.
Since you were using 2x J270 / PF5102, Idss can be as high as 20mA.
That is why in the V00, the front end has 15mA bias, resulting in 150mA for the MOSFETs.
But that is only due to the model used and one can certainly change the Vto to get the bias as you want.
In the V0 model using 2SK170/2SJ74, the bias is "normal" -- 4mA for JFETs and 26mA for MOSFETs.
But as I said before, the effect of the base resistors may just be a Spice phenomena and may not be real .....
Are you using a current limiter for the MOSFETs as in the original F5 ?
Patrick
The second stage bias is roughly determined by the drain resistor of the JFET and the source resistor of the MOSFET.
i.e. about a factor of 10.
Since you were using 2x J270 / PF5102, Idss can be as high as 20mA.
That is why in the V00, the front end has 15mA bias, resulting in 150mA for the MOSFETs.
But that is only due to the model used and one can certainly change the Vto to get the bias as you want.
In the V0 model using 2SK170/2SJ74, the bias is "normal" -- 4mA for JFETs and 26mA for MOSFETs.
But as I said before, the effect of the base resistors may just be a Spice phenomena and may not be real .....
Are you using a current limiter for the MOSFETs as in the original F5 ?
Patrick
Last edited:
I can verify the effect of the base resistor also in version V0 (with K170/J74).
Bear in mind that the simulated bandwidth of V0 is already factor of 10 lower than V00.
So one has to use a heavy hand on the Rb's.
With 100k, the effects are clearly visible.
With 1M, the base resistors are limiting bandwidth (80kHz -3dB).
Not suggesting by any means that you add 1M Rb's.
Only to demonstrate its effect, also in a slight different schematics with "normal" bias values.
I guess the first thing I would try is a Toshiba frontend .....
Patrick
Bear in mind that the simulated bandwidth of V0 is already factor of 10 lower than V00.
So one has to use a heavy hand on the Rb's.
With 100k, the effects are clearly visible.
With 1M, the base resistors are limiting bandwidth (80kHz -3dB).
Not suggesting by any means that you add 1M Rb's.
Only to demonstrate its effect, also in a slight different schematics with "normal" bias values.
I guess the first thing I would try is a Toshiba frontend .....
Patrick
In fact, the bipolars at the source resistors of the MOSFETs are the servos. They will force the current to be roughly equal, depending on the precision of the resistors and VBE matching, so its good but not superb. I was hoping for the feedback to remove most of the rest, and live with a little DC offset, which is no issue for electrostatics (but is a big issue for electrodynamics).
The built circuit actually confirms that these assumptions work. I will try and replace the input JFETs and also play with base resistors on the cascodes. But only after skiing vacation