JimL

Let’s look at the input stage in more detail – I don’t feel too guilty about putting this here since this discussion is more detailed (and longer) than the AudioXpress article – word count limitations and all that. As originally designed by Stax, it is a long-tailed cross-coupled cascode phase inverter with the right tube grid grounded. This can easily be converted to a balanced differential stage by un-grounding the right tube grid and using it for the negative input of a balanced signal. Unfortunately, if we put a signal into the left tube gird with the right tube grid grounded, we don’t get a completely balanced signal coming out. There are a number of things that can be done to improve the output balance. First, the higher the gain of the section, the closer we can get to a balanced output. A cascode has a high gain, so that helps. Second, the cross-coupling between the plate of the lower tube and the grid of the upper tube improves the balance, as explained by D.R. Birt. He describes it in an article in the June 1960 issue of Wireless World. He notes that this cross-coupling does not alter the overall gain of the cascode, thus for purposes of calculating the gain, one can neglect the cross-coupling and use a simple cascode gain calculator. In addition, since each upper tube section is being driven by both lower tube sections, the resulting push-pull action should help cancel out any even order distortion generated in the lower tubes – and the 12AT7 in particular has a significant amount of second order distortion. Third, the higher the value of the cathode resistor, the better the AC balance, thus, replacing the cathode resistor with a current sink improves the output balance. A simplified way of looking at it is, if a signal electron travels down the left input tube, when it hits the current sink (which ideally is an infinite AC resistance) it bounces over into the right input tube where it travels up. Not only is balance improved but this maximizes the fidelity of signal transmission between the two sides of the phase splitter. With an ideal current sink at the cathodes, the imbalance is approximately 1/(stage gain). With cascoded 12AT7 tubes, the stage gain is over 100, so the imbalance is less than 1%, which approaches the degree of balance using matched resistors, and significantly exceeds our ability to match tube sections – not to mention the ability to keep them matched as they age. Finally, the balance pot between the two cathodes helps static balance by allowing the output plates to be adjusted to sit at the same voltage. The capacitors connecting the cathodes bypass the pot at audio frequencies, preventing the adjustment from altering the gain from one side versus the other. Really quite a sophisticated design for something that appears so simple. Now, what are the advantages and disadvantages of this circuit? The advantages are simplicity with high gain and good output balance. There are two disadvantages, however. First, in order to achieve high gain, the output resistors are high impedance, which means that the current running through the circuit is low, so the current drive to the output stage is puny. For 300 kilohm plate resistors and +/-325 volt power supply, each tube has less than 0.55 mA current available to drive the Miller capacitance of the output stage. We will discuss Miller capacitance further in the next paragraph. Second, a cascode has a high output impedance, roughly equal to the value of the output resistor(s). Now, remember that, in AC terms, the output resistance comprises not only the plate resistor, but also the grid resistor in the following stage, which is in parallel with the plate resistor. With 300 kilohm plate resistor, 500 kilohm output grid resistor, and the output tube resistance, the output impedance of the circuit is about 173 kilohms on each side. This high output resistance is driving the output stage, which has a Miller capacitance equal to (μ+1)*(grid to plate capacitance), where μ = the gain of the output stage. For a 6SN7GTA output tube, the grid to plate capacitance is approximately 4 pf and μ = 20, for a triode-connected EL34, the grid to plate capacitance is approximately 10 pf and μ = 11, so the Miller capacitance is approximate 84 pf for a 6SN7GTA and 120 pf for a triode connected El34. Note that the Miller capacitance of the output tubes is the same order of magnitude as a set of electrostatic headphones. However, since the output tubes also produce gain, the amount of current needed to drive them is proportionately less than the current needed to drive the headphones. On the other hand, with the high output impedance of our input stage, this means that the open loop frequency response is -3 dB at 11 kHz for a 6SN7GTA output tube and -3 dB at 8 kHz for a triode-connected EL34. This circuit NEEDS overall feedback to achieve a flat frequency response over the entire audio band. That means for the entire circuit to get to 20 kHz closed loop without rolling off, it needs 6 dB open loop gain over closed loop gain with a 6SN7GTA, and 9 dB open loop gain over closed loop gain with an EL34. Now, before you throw up all over your shoes because of this crappy open loop frequency response, let me point out that the Dynaco Stereo 70, a class B Stereophile recommended amplifier back in the good old days of the late J. Gordon Holt, has an open loop frequency response that is -3 dB at around 7 kHz. Yep, 7 kHz. Still sounds pretty good. Anyway, we have two good reasons to want the input stage gain to be high: first, it maximizes the balance from an unbalanced signal going into the balanced output stage, and second, it provides the extra open loop gain needed to flatten the frequency response up to and beyond 20 kHz. Incidentally, in a cascode, the lower tube is the major determinant of the gain, however the upper tube does most of the voltage swing. Some people think a cascode is a two stage design with the lower tube as the “input” and the upper tube as the “driver.” However, in a cascode, the same signal current runs through both stages. In the usual two stage input/driver design, the input stage performs much of the voltage gain and runs at a low current, while the driver stage generally runs a higher current to “drive” the output stage. Thus, it makes no sense to use a high current tube as the upper tube of a cascode, if the circuit as a whole is running at a low current, as is the case here. Using a high current tube, say a 6BX7, as the upper cascode tube in this circuit just means that you are running it at a trickle current where it is non-linear. You may like the sound of distortion, but it is by definition malfunctioning. As a general rule, a well designed tube circuit is optimized for the tubes that are used in it. The notion that any tube that you can jam into the sockets with or without the use of a socket adapter is just hunky dory (aka tube rolling) as long as no smoke or sparks occur is, to put it bluntly, stupid, unless you like the sound of a malfunctioning, high distortion circuit if you use the wrong tube. In light of this, let us consider some candidate input tubes. Stax used 12AT7s for both input tubes in the cascode. The calculated gain using these is around 42-44 dB, depending on what parameters you use. The 12AX7 and 6SL7 do approximately as well, yielding calculated gains of around 41-43 dB. On the other hand, using 6SN7s gives a calculated gain of around 37-41 dB. These numbers don’t seem to be that much different, however remember that 3 dB represents a 41% increase in voltage gain. Also, it is important to note that the 12AX7 and 6SL7 are designed to be linear at low currents, whereas the 6SN7 tube really wants to have about 10 times as much current to be in its linear range. So using a 6SN7 in this circuit produces a non-linear result with reduced gain - like I said, a malfunctioning circuit. So for best results, the input circuit should use 12AT7 tubes, which is what it was designed for, with the 12AX7 or 6SL7 as possible acceptable substitutes. Now, as I mentioned before, since the upper tube in the cascode does most of the voltage swing, it is not unreasonable to have the 12AT7 as the lower tube, which determines the gain, and a 12AX7, 6SL7 or 5751 (roughly a miniature equivalent of the 6SL7) as the upper tube as they are designed for linear voltage operation at low currents. I confess I have not tried this as I much prefer listening to music to listening to different tubes.

You guys are clearly barking up the wrong tree.

According to the Arctic silver website, ceramique "does not contain any metal or other electrically conductive materials. It is a pure electrical insulator, neither electrically conductive nor capacitive." So it should be OK.

I submitted the article again last month - apparently they didn't get it the first time. Haven't received word on when the plan to publish, but I believe it's going to come out in two parts, with the power supply in the second part. It's a pretty basic supply, derived from two articles in TubeCad but using MOSFETs instead of tubes, plus a stabilized TL431.

Re Arctic silver 5 - from the website: "Made with 99.9% pure silver: Arctic Silver 5 uses three unique shapes and sizes of pure silver particles..." Siver is definitely conductive. Maybe not mixed in a non-conductive paste for heatsinking in a computer where the highest voltage is 12 volts, but definitely possible when the voltage is 450 volts and up! Air isn't conductive either - but shuffle across a rug and touch a metal doorknob...and air doesn't contain micronized silver.

Gary Pimm discusses using the mu output in the section on constant current sources on his website. Basically, this involves taking the output signal off the junction between the lower and upper device in the cascode current source. In the cascode, the upper device takes nearly all of the voltage variation, isolating the lower device, which runs at very close to constant voltage, and hence constant current. This means the output tube also runs at constant current, which is its lowest distortion mode, and the output signal current comes from the upper device rather than from the output tube. Since the lower device is running at a constant voltage, like a battery, the mu voltage output closely follows the plate voltage. The upper device is low impedance and so able to drive large capacitative loads better than the tube plate could. However, as Pimm points out, if you take the signal current from the tube plate which is loaded by a constant current source, the power supply is only responsible for providing a pure DC current to the circuit, and the signal current only circulates in the active circuit. The current sources isolate the active circuit and the signal current from the power supply. If you take the signal current from the mu output, since the output tube is now running at a constant current, the signal current must come from the power supply, which means that the supply, with its non-ideal electrolytic capacitors, is now in the signal pathway. This is a significant disadvantage. OTOH, the capacitative load of a stat headphone is quantitatively similar to that of a grid resistor and output tube, so the lower impedance and improved drive capability of the mu output does not provide a significant benefit in exchange. This can be partially alleviated by using a shunt regulator fed by a constant current source. This would remove the passive power supply capacitors prior the regulator current source from the signal pathway, however the shunt regulator itself would still be in the signal pathway. In the SRX Plus, my goal was to maintain the integrity of the original circuit while optimizing its function. The best way to do that was to replace the output resistors in the original circuit with constant current loads. The additional benefit was isolating the active circuit from the power supply.

Although I said there was no way to balance output tube sections in a previous post, I realized recently that I was mistaken. In the amp I built there is no way to adjust because I built the output loads with fixed resistors, however since the published schematic has adjustments in the output current loads, this can in fact be used to vary the cathode-to-plate voltage individually for each output section by slightly altering the current running through each section. The concept is, start with the output current loads fixed, adjust the cathode current sink to approximately zero the output plates, then vary the current loads to adjust the offset between + and - voltages for each channel. The three adjustments are interactive, so, it would probably be best to adjust one of the current loads to partially decrease the offset, the adjust the other in the opposite direction to further decrease the offset until that is balanced, then adjust the current sink to zero both plates.

The main point of my previous calculations is to show that a constant current load is a much better way to drive a stat headphone than a resistor (or a plate inductor for that matter), because using a really good current load basically removes it from the equation - all the driving triode sees is the headphone load. Much better efficiency in terms of getting the signal current to the headphones. So for example, the SRX Plus can do about 2/3rds the output for the same ratio of signal current to standing current as the BHSE, with less than 25% of the power in the output stage. A good big 'un can beat a good little 'un but the little one can come surprisingly close by working smarter, not harder. The problem with a pentode or cascode as an output stage is its high impedance, which means it doesn't vary its current much as the output voltage changes. This is what we DON'T want in an output stage. For example, if we were to combine a cascode or pentode with constant current load, the output voltage could vary all over the place (high voltage gain) but the output signal current would be very low. Think of the reductio ad absurdum where the cascade or pentode is modeled as a constant current source, into a constant current load. The voltage would move up and down but there would be no signal current output, and you need both signal voltage and signal current to generate power, By the law of conservation of energy, no power, no sound output As the old saying goes, "it couldn't drive its way out of a wet paper bag." Now, the problem of using a resistor load for the cascode/pentode as the driver is that you need to have a low impedance resistor so the current can easily go to the headphones. So, let's pick a 5 kilohm resistor for the plate of our output stage to provide a sufficiently low impedance driver for the headphones. If we want to swing 400 volts peak, using Ohms law, we have to modulate the current in the resistor by 80 mA, which means we need at least 80 mA standing current in the resistor, which is 32 watts of power - we need a BIG resistor, and also a BIG tube because that also has to burn up 32 watts. Now, if we have a differential output stage that is 128 watts/channel power dissipation, not counting the heaters for the tubes. Note this is to swing a mere 800 volts peak-to-peak, if we want to swing 1200 volts P-T-P increase everything by 50%. Kinda makes a BHSE with its 40 watts/channel standing power dissipation seem like a paragon of energy efficiency. And since the resistor is so much lower in impedance than the headphones, 99.8+% of the signal current is wasted in the resistors. So, a long-winded way of saying, not a good idea.

dsavitsk: So if you have a current source on a cascode or pentode stage the benefit is very high gain, but the down side is very high output impedance. Loading it with a resistor will decrease the gain, and loading it with a tube, with its Miller capacitance, will result in rolled-off frequency response. This is why I didn't put current sources on the input stage cascode of the SRX - with the 300 kilohm resistors in the basic circuit, the open frequency response rolled off with a -3dB at around 11 kHz, and substituting a current source would make things worse. For an amp output stage you need to have a relatively low output impedance to supply signal current to the transducer, whether it be a speaker or a headphone. Most speakers are designed to be driven by voltage sources, so most speaker amplifiers have an output impedance of 1 ohm or less, often very close to zero. Now, a triode connected EL34 such as in the ESX has a plate resistance of around 1 kilohm, and a 6SN7GTA such as in the SR Plus has a plate resistance around 7 kilohms without feedback. Feedback does help, cutting the effective output impedance down to around 400 ohms for the EL34 in the ESX and around 1400 ohms for the SRX Plus. That sounds fairly high, but remember that a 100 pf cap has an impedance around 80 kilohms at 20 kHz, which rises at lower frequencies. Laowei: Good point. With the exception of a couple designs which have no neg feedback, most stat amps have feedback resistors around 200k to 250k, which with 400 volts peak amounts to another 1.6 to 2 mA required from the output devices. So now, including the current going to the NFB loop we have: ESX 50k load resistors: 12.6 mA = 79% of standing current ESX 10M90S current source: 7.0 mA = 44% of 16 mA standing current = 25% of 28 mA standing current = 19% of 36 mA standing current SRX Plus cascode CCS: 4.4 mA = 31% of standing current

Wow, fantastic build, Kerry!

Let me get back to something I discussed briefly in my first post but haven’t addressed since then, and that is that stat headphones must consume power because they produce sound. Now, all the calculations I’ve done up to now are based on a simple model of the electrostatic headphone as a capacitor. However, even though capacitors use require current from the amplifier to swing voltage, they don’t use up power because the voltage and current are out of phase. And of course, we don’t listen to capacitors because they don’t make a sound (at least in theory). Since headphones do make sound, that means they must use extra current from the amplifier beyond what we have previously calculated. So let’s take our previous example of 800 volts peak-to-peak. This is 400 volts peak, and since Dr. Gilmore has stated that the SR007 can consume one watt at peak output, let’s say the peak current is 2.5 mA, which works out to that one watt peak. If we add this current demand to our previous results, we get: ESX 50k load resistors: 10.6 mA = 66% of 16 mA standing current ESX 10M90S current source: 5.1 mA = 32% of 16 mA standing current = 18% of 28 mA standing current (T2) = 14% of 36 mA standing current (BHSE) SRX Plus cascode current source: 2.9 mA < 14% of 14 mA standing current Again, these are “back of the envelope” calculations so don’t take the absolute numbers too seriously but the relative relationships should hold good. We can now see why pongo5's Egmont sounds much better with constant current source loads. The Egmont with 60k load resistors is similar to the ESX with load resistors in terms of current demand, whereas the Egmont with cascode current sources is similar to the SRX Plus. The basic circuit draws 3x as much signal current for the same signal voltage because the load resistors waste so much of it swinging the voltage in the resistor.

Either you were lucky or I was unlucky or both. I found 5-10% deviations in current with the same source resistor was fairly common, but I also ordered various DN2540 over a period of time. Also, I was trying to get within 1% of designed current so my OCD is probably showing.

With those colors, I bet if you stared at one side of the board for 30 seconds, then shifted down to the other side, you'd see the whole board layout at once. More seriously, one concern I have is that while the 10M90S/DN2540 makes for a very high impedance CCS (good) there is a fair amount of variability between DN2540 samples in the amount of current that is produced for a fixed source resistance (bad). That's why I put a trimpot in the source to adjust the current. This is unlike the 10M90s which is designed as a current source so is reasonably tightly controlled for current vs source resistor.

Actually, it looks like a KGST with a solid-state equivalent of the BH output stage (grounded grid/gate) and cascode current sources, if I'm not mistaken.

Thanks for the comment, Dr. Gilmore, that's a very good point. One thing about the cascode CCS (constant current source) on my SRX Plus (modded SRX) is the output voltage drifts about 40-50 volts as it warms up, also the final value may vary by a few volts with different turn-on/turn-off cycles, presumably due to thermal issues. Thermal stability is very significant for practical amplifiers. In terms of the LL, my copy of the prototype schematic courtesy of Dewey, Cheatem and Howe , uses the same topology as figure 3A discussed in my previous post, but substitutes a MOSFET for a BJT as the device that has the large voltage swings in the output CCS. Here is what Jung has to say a few paragraphs earlier in his introductory remarks in the section discussing the one VBE current source shown in figure 3A: "Of course, higher-voltage parts should be used when appropriate. While exotic and super-high-gain parts aren't necessary for very good performance from these circuits, LOW CAPACITANCE DEVICES DEFINITELY ARE PREFERRED (<10pf), A CRITICAL POINT IF SUBSTITUTING [emphasis added]." The reason is that high capacitance devices cause the effective impedance to decrease at high frequencies, degrading the performance of the current source. Given that MOSFETs have a relatively high shunt capacitance, this means that performance of the LL output current source is likely to be even worse than the BJT version that Jung cited as a design to avoid. Now the "fixed" version of the LL uses a low-capacitance BJT for the MOSFET, which is better, but.... the SRM323, which shares a very similar topology, uses the LED and BJT CCS on the output, which is significantly better than the CCS of the "fixed LL." You can't say that Dr. Gilmore and spritzer didn't warn us about this when the LL came out, although they didn't go into it in the infinite gory detail that I just did. Now, one of the advantages of a cascode MOSFET CCS such as the SRX Plus uses, is that the upper device shields the lower device from voltage variations, so that the lower device, which sets the current, sees nearly constant voltage regardless of the voltage variations across the cascode pair. This means that the effective capacitance of the cascode is very low, so its performance is preserved to high frequencies. For example, Pimm measured a cascode DN2540 pair as having an effective capacitance of < 0.2 pf! By comparison a single DN2540 measured about 32 pf. ADDENDUM: I got a look at the production schematic courtesy of spritzer's LL Mk I thread elsewhere on this site. The output current source is identical to the Dewey, Cheatem and Howe version with the exception of a protection zener added to the MOSFET in the production version, so same lousy performance.

Those who are interested in finding out more about constant current sources should read the two part articles by Walt Jung which were published in AudioXpress in 2007, with addendum in 2009. The first and second parts are available on the web, just do a search for Walt Jung, the title is: "Sources 101: Audio Current Regulator Tests for High Performance", which shows a number of different current source designs with measurements of performance. Interestingly, while browsing through the first part, I found (Figure 3A) a low voltage version of a design that is identical to the constant current source used in the output section of the eXstatA, except with high voltage NPN transistors instead of low voltage PNP transistors that Jung used in the article. Jung measured its dynamic impedance as roughly equivalent to a 45 kilohm resistor. He comments that "the performance of this circuit is easily bettered by many others, both more simple and cheaper. For these reasons, plus the stability caveat, this circuit isn't recommended. Except as an example to avoid, perhaps." If the high voltage eXstatA design is no better then it is not clear the advantage that 6 parts plus heatsink has over a high power resistor. By contrast, the low voltage version of the transistor plus LED current source that is commonly used in Dr. Gilmore's designs, for example, has a measured dynamic impedance of around 170-180 kilohms which is similar to the results with a single 10M90S current source. One can do even better substituting a reference diode such as an LM336Z2.5 which bumps the dynamic impedance up to around 3 megohms, although Jung didn't address the noise performance of the substitution. Anyway, for those interested in the nitty-gritty of design, well worth reading IMHO.

Doh! I deleted my last post because it was totally wrong. Sorry about that. So, the problem right now is that the cathode-to-plate voltage is too low, thus the output voltage is negative with respect to ground. In order to increase the cathode-to-plate voltage we need to increase the cathode-to-grid voltage difference. You can see this if you look at the tube curves, going to a more negative grid voltage along a constant current line increases the plate voltage. In order to increase the cathode to grid voltage difference we have to increase the cathode resistor. Since we are using current sources with a total of 10 mA into the cathode resistor, the calculation is relatively easy. A 100 ohm change in the cathode resistor will result in a 1 volt change in the cathode-to-grid voltage. Looking at the tube curves for the 5965, as we run along the 5 mA current line, decreasing the grid voltage by one volt will result in about 30-40 volts increase in the cathode-to-plate voltage. So, for each channel, take the average offset between the + and - outputs, and substitute a cathode resistor that is more by the calculated amount. One caveat, the offset needs to be checked when the amp is warmed up with the cover on. Oh, and good job, pongo5! It's always gratifying when practice confirms theory.

I love the smell of burning fiberglass in the morning. It smells like....Kingsound!

On the amp board there are two sets of four 12k resistors, which I'm thinking are the plate resistors in series connection to make 48k. Those resistor values don't make sense as grid resistors and probably not as cathode resistors in parallel either, as the cathode shouldn't need that much power handling, so what else could they be? Not sure of the values or function of the two large resistors between the red capacitors. Anyway, if I'm correct, with +/- 300 volt supplies and the tubes biased so that half the voltage is across the plate resistors and half across the tube, that would make the standing current about 6 mA/plate and the tube dissipation about 80% of max.

Overall view looks like PS board at left, unbalanced to balanced board at top and two amp boards. I'm guessing the unbalanced to balanced board uses a 12AT7 as a long tailed phase splitter with a 2SK170 current source on the tai to improve output balance, with capacitor coupled outputs. A top view of the amp board before assembly would be helpful. Looks like a 12AT7 driver capacitor coupled to a 12AU7 output tube, also capacitor coupled. Load resistors on the outputs which means a lot of the signal current burned up in the load resistors. Not an A10 clone - not enough tubes for SRPP. I'm thinking more parts so more complicated than an SRX, but less sophisticated. In terms of potential output power based on combined plate dissipation, 12AU7 < 6CG7 (used in SRM006) < 6SN7GTA (modified SRX) < 6S4A (KGST) < EL34 (BHSE). The SRM007 uses 6CG7 in parallel so more potential output power than a 6SN7GTA but not as much voltage swing, however the 007 uses load resistors so using good current sources with a 6SN7GTA puts it back on top in terms of available output power.

Wow. I could call Mikhail a clueless fucking idiot . . . . but I'd hate to insult the clueless fucking idiots out there.

So to elaborate on spritzer's comment, the 6S4A is a single triode tube, the 12AU7 is a dual triode tube. The 6S4A triode has a much higher max plate voltage (550 volts vs 330 volts) and power dissipation rating (8.5 watts vs 2.75 watts), different filament requirements, they are wired completely differently. their internals look completely different, and neither will work in a circuit designed for the other. They do plug into the same socket so they must otherwise be completely identical . In summary, they are as alike as Arnold Swarzenegger and the Bobbsey twins. If your friend can't tell the difference between the two tubes he likely can't tell the difference between Arnie and Bobbsey.

Another example, let's take the ES-X but with 10M90S current sources instead of the 50 kilohm resistor load. I measured the dynamic impedance of a single 10M90S current source at about 170 kilohms, so for 800 volts peak-to-peak as in the second post, the worst case current demand is about 2.8 mA, about 17.5% of the total idle current of 16 mA/channel. This is much better than the >50% demand using the resistor load. But now, we can see another benefit of using current loads. There is no reason why we need to keep the idle current at 16 mA. We can set the idle current at any level we wish, within the limits of tube dissipation and heatsink size. For example, if we increase the idle current to 28 mA (like the T2) or 36 mA (like the BHSE), the worst case current draw is only 10% (T2) or < 8% (BHSE) of the idle current. Aside from increasing the current reserve, which should decrease distortion, why would we want to do this? Well, running the EL34 at 8 mA per tube is really a pretty low current, where the tube linearity is not as good as it could be. Running the tube at a higher current puts it into more linear region, further decreasing distortion.

A cheap CANADIAN clone? Who woulda guessed?