Now look at the currents "stolen" from the ~10mA CCS for either case at 20kHz and same load, to drive the MOSFETs, which one do you prefer (look at the magnitudes)? 15uA or 2mA error current sure make a difference for a 10mA CCS in terms of nonlinear error magnitude. To unload the high impedance node is always preferable for controlled conditions, you still can load it deliberately, resistive to lower DC gain and/or capacitive to shift the pole to where you need it.Lumba Ogir said:
- Klaus
Attachments
Lumba Ogir said:
I think Bob Cordell wouldn't agree with that. Anyways my simulations with the -extra- driver make the amp very unstable, which is not surprising as the hidden compensation scheme made by the gate capacitances changes doing so.
Mosfets requiere a single driver but we are already providing this, well, in fact the driver is also a hidden VAS with 100% feedback, but there is no problem in seeing it as a driver.
KSTR, your plots look uglier than mine, are you sure that you are running the mosfet hot enough?
EDIT: I have just noticed that your plots show hysteresis. I can't see a single reason why it should be so. I did read that one of the IRFP IR mosfets (i don't know if it was the P one or the N one). IR constantly says that these aren't suited to linear. I think that the post talking about this phenomenon said that it did not affect the fairchild ones, which also state that they have "good linear characteristics". I think i will buy some fairchild ones when i order the panasonic 105º FC caps that i use in class A amps (I choose expensive caps to be sure they will be able to cope with the high temperatures).
Lumba Ogir said:
I don't agree with you in that mosfets should be direct-driven in a standard 3-stage amp. They need a driver in that context, but in this amp they don't need it because the VAS is not a vas. Sorry, i didn't realize it. The VAS is the output mosfet. The driver is the differencial transconductance input stage. It translates the voltage difference between input and output into current. The output stage works as an integrator and translates current (charge) into voltage like a vas does after the pole, which has nothing to do with the gate stopper because the output impedance of the driver is very high if not cascoded and immense if cascoded . Sorry for the previous mistake. The output mosfet does not work in unity gain. A driver changes the transconductance stage + integrator topology and would only work if the driver acted as a VAS. This would requiere extra compensation and probably would be very difficult to make stable.
BTW i have taken a 2SJ162 model from a thread here and i found the simulation results in that amp, well, hard to belive. THD 0.025% full output at 25KHz is fine for a class A amp, but THD of 0.07 % at 10 MHz is suspicious. Anyways i will buy some of these... just to try. The unquestionable advantage will be the higher efficiency compared to the IRFP's.
EDIT: the driver can be turned into a vas by degenerating the CCS.
EDIT2: If a driver is added the system will behave in a way much more similar to standard amps. In the non-cascoded direct-driven amp the first transistor stops behaving as a voltage gain device and becomes a transconductance device is arround 1KHz. Cascoding sends this to the subbass region. Adding a driver sends this to ultrasound region. In total agreement with Hugh this deserves listening tests.
Hi Ion,
If you support a CCS from the positive rail to supply the npn driver with current, it will require about 1.6 volts of headroom for proper function. This will limit max peak output, of course. So the mosfet + degeneration at quiescent will need at least this 1.6V of bias; this is not possible with the mosfets you mention, only with hexfets like IRFP9140 which have much higher transconductance anyway. I recommend the commonly available IRF devices since at 1.4 amps they bias at around 3.5 volts which leaves plenty of room for the CCS. Source degen on these devices pulls back the transconductance to 1-2 siemens, linearising output without much penalty to Zout.
Your description of operation was interesting.
The driver is operating at constant current, therefore it is not, as I see it, a transconductance amplifier but a form of VAS with less than unity gain. It drives the mosfet gate with voltage only at constant current, and not too much of it; the driver emitter moves up and down with input signal, say 15Vpp, yet the collector hovers close to the positive rail and moves from 3.5V below the rail to perhaps 5V below the rail at the most. The gain of this driver would be given by the usual ratio of collector to emitter impedances; the former would ultimately be limited by the input capacitance of the mosfet, not the CCS, while the emitter resistance would essentially be 26/mA of collector current (I believe we can safely discount the speaker load because its effect is bootstrapped by negative feedback so the mosfet 'buffers' the driver). This reveals a very high open loop gain and consequently high fb factor, verified by the very low Zout of the Sziklai, measured at 80 milliohms for small signal into 8r resistive. This concertina operation is interesting to watch on a CRO, but it is still a voltage drive to the gate.
The mosfet, OTOH, is operating as a transconductance amplifier since there is voltage in and current out. But there is also a voltage change at the drain forced by the driver, just like the plate of a tube.
You've spoken of cascodes; one cascode I like is supporting a bipolar npn cascoding device above the driver holding constant voltage across the driver. A jfet driver can then be used, giving very high input impedance to the stage. This places the driver in constant power so it's Vbe cannot change, and since the transfer function of this amplifier is essentially a step function, comprising the Vbe of the driver, then the linearity of the entire amplifier focusses on the variation, however slight, of the Vbe of the driver. This approach can also be used on sziklai pairs with gain.
My email is aspen1 at people.net.au if you wish to discuss this further.
Cheers,
Hugh
If you support a CCS from the positive rail to supply the npn driver with current, it will require about 1.6 volts of headroom for proper function. This will limit max peak output, of course. So the mosfet + degeneration at quiescent will need at least this 1.6V of bias; this is not possible with the mosfets you mention, only with hexfets like IRFP9140 which have much higher transconductance anyway. I recommend the commonly available IRF devices since at 1.4 amps they bias at around 3.5 volts which leaves plenty of room for the CCS. Source degen on these devices pulls back the transconductance to 1-2 siemens, linearising output without much penalty to Zout.
Your description of operation was interesting.
The driver is operating at constant current, therefore it is not, as I see it, a transconductance amplifier but a form of VAS with less than unity gain. It drives the mosfet gate with voltage only at constant current, and not too much of it; the driver emitter moves up and down with input signal, say 15Vpp, yet the collector hovers close to the positive rail and moves from 3.5V below the rail to perhaps 5V below the rail at the most. The gain of this driver would be given by the usual ratio of collector to emitter impedances; the former would ultimately be limited by the input capacitance of the mosfet, not the CCS, while the emitter resistance would essentially be 26/mA of collector current (I believe we can safely discount the speaker load because its effect is bootstrapped by negative feedback so the mosfet 'buffers' the driver). This reveals a very high open loop gain and consequently high fb factor, verified by the very low Zout of the Sziklai, measured at 80 milliohms for small signal into 8r resistive. This concertina operation is interesting to watch on a CRO, but it is still a voltage drive to the gate.
The mosfet, OTOH, is operating as a transconductance amplifier since there is voltage in and current out. But there is also a voltage change at the drain forced by the driver, just like the plate of a tube.
You've spoken of cascodes; one cascode I like is supporting a bipolar npn cascoding device above the driver holding constant voltage across the driver. A jfet driver can then be used, giving very high input impedance to the stage. This places the driver in constant power so it's Vbe cannot change, and since the transfer function of this amplifier is essentially a step function, comprising the Vbe of the driver, then the linearity of the entire amplifier focusses on the variation, however slight, of the Vbe of the driver. This approach can also be used on sziklai pairs with gain.
My email is aspen1 at people.net.au if you wish to discuss this further.
Cheers,
Hugh
AKSA said:
We are focusing on different parts of the whole picture. There are two situations. At low frequencies where the gate impedance is very high the system behaves as you described, but when the impedance becomes lower with frequency (in comparison to the collector-emmiter resistance of the driver) the system starts working as a transconductance device. The changes in voltage between the input and the output make the current change through the driver, and this current can't go into the ccs so it gets stuffed (integrated) into the gate of the driver. The pole i'm talking about is the frequency at which the system changes its behaviour.
Cascoding raises the "early" collector to emmiter resistance of the driver and makes this frequency lower as the output capacitance is left unchanged while the driver resistance is increased, so the point where they meet is at lower frequency.
Adding a buffer before the gate makes the output impedance look bigger (roughly multiplied by beta) so the frequency where they are identical (from the driver's point of view) becomes higher. This can also be seen as the gate seeing a lower driver impedance instead of the driver seeing a higher gate impedance, but the consequences are obviously the same.
What makes me think there is room for improovement is that this frequency is inside the audio band on the non-cascoded non-buffered version, and this may introduce a change in the character of the amp. Being this at (roughly) 1KHz it might "split the midrange". No idea if the effect is actually audible, but the circuit changes its behaviour completely there.
I have jsyt done a quick sim to see if there was a change between the distortion at 500Hz and 5KHz and it isn't there, so i might be wrong.
Ion,
Triangulation of a square wave, always a good indication of poles, lies around 42KHz IIRC. There is rounding of the square wave at 10KHz to around 25% into the horizontal.
This would indicate that the pole is probably around the 20KHz mark. I would expect it to relate to the CCS current and the gate parasitics, let us assume 700pF.
Did you receive my email?
Ogir,
Please, it is not nonsensical, it is merely another man's way of looking at things. Who knows, Ion might be right!
Hugh
Triangulation of a square wave, always a good indication of poles, lies around 42KHz IIRC. There is rounding of the square wave at 10KHz to around 25% into the horizontal.
This would indicate that the pole is probably around the 20KHz mark. I would expect it to relate to the CCS current and the gate parasitics, let us assume 700pF.
Did you receive my email?
Ogir,
Please, it is not nonsensical, it is merely another man's way of looking at things. Who knows, Ion might be right!
Hugh
Lumba Ogir said:
I think you are grossly overestimating my knowledge. Minimizing Cgs can be archieved by cascoding, but this forces to add a compensation capacitor of roughly the same size (multiplied by the gain) to get the same stability. To me it seems a no-go unless there is another way to make it stable.
AKSA said:
Yes, and i did reply, have you received it?
I assume the gate parasitics to be much bigger, arround 10 nF. That's because the Cgs is not bootstraped by the load and the Cgd is magnified (the voltage between gate and drain raises what the driver raises it plus what the output goes down, and thus requieres arround 17 times more charge ( the current increases gm per volt (2A per volt) and the loudspeaker voltage goes 8V lower per A, so the voltage increases 1V + 16V per volt and the charge needed is the same as if Cgd was multiplied by 17)
EDIT: Sorry about my weird way of explaining things. Lumba Orgir is partially right, but at the moment i can't do better.
Hi Ion,ionomolo said:
The mosfets (2 per leg, 4 in total) run right at their typical safe linear power dissipation of ~30 Watts, with ~30V that's 1A each. The "raggedness" in the waveforms comes from that change of Cdg vs. Vds. When I load the output with current only (zero voltage swing) I get the expected quadrature relation between mosfet gate and drain currents. This again would be the point where cascoding the mosfet could make a difference, higher loading though, but more linear, more true "cap-like" (unless it goes into cutoff). OTOH when Vds is high enough, then Cdg is low and quite constant, therefore the problem is mostly relevant under high voltage swing close to the rails, not the usual situation with music.
In the lower plot there also is the effect of the sliding bias which already kicked in which additionaly distorts the waveforms.
I've been using the N-Ch IRFP244 just because I had in the back of my head the known problems with the P-Ch part's transconductance drop with higher frequency (but that most probably isn't modelled in any IRF9xxx model anyway).
"Less suited for linear", when used in conjunction with vertical MOSFETs designed for switching, usually further refers to the hotspotting problems in linear operation mode of high gm parts with their strongly positive tempco, the bigger the die the more severe the problem (and the reason for the low safe linear SOA despite high Pd_max ratings in the datasheets).
- Klaus
KSTR said:
I remember to have read that the hysteresys problem i'm talking about happened at LF. You are saying that the hexfet mosfets show someting similar to secondary breakdown, i also remember having read about it, but general experience is that they last forever in a circuit as long as they are kept inside the safe operating area, but this means big heatsinks, of course.
- Status
- This old topic is closed. If you want to reopen this topic, contact a moderator using the "Report Post" button.