One of things I didn't like about my orignal SOHA design was the trimpot. This problem is solved nicely in the SOHA II with an input stage that permits a small amount of feedback which, in the SOHA II, really helps with tube rolling and gives excellent performance.
However, in this amp we don't have room for the nice front end in the SOHA II that needs two tubes. We can only have one tube.
So how do we set the plate voltage for the tube without loading it down too much (destroying the high frequency response)? First, let's take another look at the SOHA front end:
The triode is in a grounded cathode configuration. In order to set Vp we simply adjust the cathode resistor using a trimpot to set Vk. This changes the bias conditions on the grid until we get the desired plate voltage. If we can find a way to auto-adjust Vk by sensing the plate voltage we can eliminate the trimpot.
Since the plate is loaded with a CCS the effect of the trimpot is to change the cathode voltage for a constant Ip. Another way to adjust Vk, however, would be to leave the resistance a constant value and either sink or source current from it. Changing the current in the resistor will effectively adjust Vk. If we can create a current source that senses Vp and knows what value to set Vp to this scheme will work. In other words, we need a current source that has a reference voltage, Vr, that it can use to compare Vp to and then adjust the current in the cathode resistor. This idea is illustrated here:
Notice that because the cathode resistor is bypassed the current source will be shielded from AC variations in Vk during normal operation.
This particular circuit feature begs to use an opamp for the current source. The opamp can source or sink current from the cathode resistor as needed. We just need to set Vr and to find a way to sense Vp. Sensing Vpis tricky because in order to sense Vp we have to attach something to the plate. When we attach anyting to the plate we immediately add a load to the tube. If we attach a load that is too small the tube's performance will suffer at high frequencies. Thus, if we use the most straightforward sensing circuit, a voltage divider, the resistors must be large - in the MΩ range.
With a voltage divider to sense Vp and an opamp current source the circuit looks like this:
The value of Vr is somewhat arbitrary needing only to fall within the opamp rails. But we can add another condition which is that most of the cathode current should flow through the resistor so that the opamp is not supplying the major portion of the current. Even this condition is arbitrary. We must make some additional choices. Anticipating the design of the PS we can expect to have Vp=80V and CRD=2mA (1N5305). Looking at plate curves for the 12AU7 running at 80V and 2mA we guess Vk about 2.5V. And knowing that the buffers will be powered by V+=24V we can make things simple and set Vr to one tenth of V+ or 2.4V so that the opamp's input and output voltages are nearly the same. Normally, we would use a zener or other reference device for setting Vr, but since the V+ will already be regulated we can use a simple resistive voltage divider.
Now, to get 2.4V from 24V we need a one tenth voltage divider. I like to set dividers to run at 1mA when possible (arbitrary). The closest 1/10 divider will be to use 18kΩ/2kΩ for the reference divider from 24V.
For the Vp sensing divider we can also arbitrarily set Rp1 at 1MΩ. It remains to calculate Rp2. The opamp will try to equalize its input voltages and, because it's a jfet opamp, will draw essentially no current. This means that Vpd will also be 2.4V. With this final value we can compute Rp2:
Vp = Vpd * (Rp1 + Rp2) / Rp2
Putting in the values we get:
80V = 2.4V * (1MΩ + Rp2) / Rp2
Solving we get Rp2=30.9kΩ
Which value just happens to be a standard value for Vishay RN55D resistors. And the final servo circuit now looks like this:
Additionally, the 1MΩ resistor is large enough not to steal any plate current from the CRD. Thus, nearly all of the CRD current becomes Ip. The opamp has a rail cap to prevent oscillation and an integrating cap to make it operate like a servo.
Using an off-the-shelf walwart necessitated some important design constraints. And working within 75mm x 120mm necessitated a few others.
Some hybrids use DC walwarts. This is the simplest way to supply DC power to the amp but it tends to limit the B+ to something less than what the walwart can supply.
However, if we use an AC walwart, as is used in other hybrids, then we have the opportunity to use voltage multipliers to create a higher B+. How high depends on the voltage from the walwart and the number of stages in the multiplier.
In this case I had settled on a 24VAC walwart as the starting point. There is an additonal limitation imposed by using a walwart. Since the walwart only supplies two wires there is no CT. And with no CT we cannot create a split supply for a complementary output stage. We could use an SE output stage as in the SOHA II but the power requirements for this are too high.
So, the two constraints of small space and AC walwart power require that the O/P stage be:
- Use TO92 size devices
- Use a rail splitter to avoid using a large output capacitor
- Minimize number of components
Number 3 further means that, since the O/P is direct coupled to the headphones, it has to be a self-zeroing buffer. This is because the input of the buffer will be coupled to the tube stage by a small coupling cap.
The simplest BJT complementary buffer is the diamond buffer with TO92 BJTs. BC327/337 complementary pair are easily obtainable O/P devices and BC550/BC560 do the same for the input pair. The operating point is should be about 20mA in the output stage. Since each device sees about 12V this makes for about 240mW power dissipation for these 600mW devices.
To minimize any possible problems due to hfe mismatch the devices should be all chosen from the same hfe class. For example, BC550/560 should be "C" class and the BC327/337 should be "25" class. Transistors shouldn't have to be matched, but class selection is a reasonable requirement.
Then next simplest improvement to the simple diamond buffer is to load the input followers with CCSs on their emitters. This helps to make the input followers more linear with consquently less distortion.
This techique is used in the PPA buffer except that the CCSs are formed using current mirrors that are themselves sourced with a single jfet CCS.
The CCSs, because they are active devices, now offer us a way to introduce a DC offset servo. If we leave the upper CCS fixed but make the lower CCS variable we can use the bottom CCS's current adjustment to set the DC offset. Note that we don't have a bias adjustment in this buffer. We must use a fixed bias because there is no room for a variable one. A fixed bias will change somewhat with the actual devices hence an additional need to ues BJTs from the same hfe class.
What we need now is an adjustable CCS at the bottom that is referenced to 0VDC. We can do this with a current mirror driven by an opamp. The opamp will be the current source and the mirror will reflect this current to the bottom input BJT. Like this:
The value for Roff requires some thinking. The opamp offset servo only controls half of the input stage. The other half is fixed by the CRD. The input BJTs have their bases in series so their base currents must be identical. Thus, if they have hfe differences their collector currents will differ by the hfe differences. This cannot be otherwise because the base currents are forced to be the same. There are two regimes that determine the value of Roff.
Qp(hfe) > Qn(hfe)
If the bottom transistor has lower hfe then its collector current will be smaller than the upper transistor. To adjust for this the opamp will try to reduce the current in the mirror by lowering its output voltage towards the bottom rail. The opamp can drive its output almost to the rail so it can reduce the current as much as it needs to to compensate. In practice it can’t do this for any possible disparity in hfe, but it can do it for anything reasonable when the lower transistor has lower hfe than the upper one. In this case, therefore, the value of Roffis not particularly critical. In practice there will be a lower limit on this resistor value because the opamp's output cannot be connected directly to the mirror. If it is, there is not much resistance there to limit the O/P current from the opamp.
Qp(hfe) < Qn(hfe)
There can be a problem, however, when the top device has lower hfe. In this case the top transistor’s collector current will be smaller than the bottom transistor. But, in order for the circuit to work the top transistor, by design, must have a collector current equal to the CRD current. This means that the top transistor’s base current has to go higher (because it has lower hfe) and this means that the collector current in the lower transistor must go higher by the ratio of the hfe of the two devices.
Now the opamp will try to drive the mirror current higher by raising its output voltage. It is limited by the top rail. The maximum current that can flow into the mirror is approximately;
Im = (V+ - V- - 1.8V - Vem) / Roff
The 1.8V is for the two diode drops in the bases of the mirror transistors and for the fact that the opamp can only get to approximately one diode drop of its rails (three diode drops total). Vem is the drop across the emitter resistor of the mirror device.
Vem = Rem * Im
Substituting Vem in the first equation and solving for Im gives:
Im = (V+ - V- - 1.8V) / (Roff + Rem)
And solving this for Roff gives:
Roff = ( (V+ - V- - 1.8V) / Im ) - Rem
To go any farther we must make a few choices. First we choose a typical value for Rem = 220Ω and for the CRD current we pick 4.3mA (1N5313) to keep the input devices in a nice class A regime.
If we want the servo to be able to handle a 3X difference in hfe (more than the spread in the "C" gain class which is 400-800) then the mirror must be able to provide 12.9mA collector current to the bottom device. If the rail-to-rail voltage is 24V then we have:
Roff = (23.2V / 16mA) - 220Ω = 1578Ω
To give some additional headroom we select Roff=1.5kΩ.
The Diamond Splitter
Because the rail splitter sources and sinks the currents in the O/P stages, the "SQ" of the splitter is almost as important as the SQ of the buffers themselves. Thus, a simple opamp-based splitter is not good enough to complement the O/P buffers. So, what do we do?
The first step is to review a few rail splitter configurations. A really good resource for designing rail splitters is Tangent's Virtual Ground Circuits. Two basic rail splitters can be made from an off-the-shelf TLE device. In this case the TLE2426 is the right one because the V+ is 24V.
The TLE is very good as a low current virtual ground circuit, but its maximum current is 20mA. Since our buffers are each running at 20mA and could possibly draw more than 40mA from the splitter we need a higher current output virtual ground. A common way to do this is to use a TLE device as only a reference for a high current buffer as shown in the right side drawing. There are many types of buffers that can be used here including single chip buffers (BUF634), high current opamps in unity gain mode, and even discrete buffers. For this amp we want to use a discrete buffer similar to the O/P buffers. The splitter butter, therefore, should have a similar a topology as we can make. In other words the splitter buffer should look like a diamond buffer. Like this:
The only thing that's different from the O/P buffers is that we don't need CCSs on the emitters and we have paralleled O/P transistors to handle the currents from both O/P buffers at the same time. All we need now is the feedback circuit that will hold the VG of the diamond splitter. To do this effectively we need lots of gain, such as, an opamp. In this case we want to include the buffer in the feedback loop of the opamp so that the opamp will use all of its gain to hold the buffer's output to the reference voltage. And for simplicity we use the TLE device to provide the reference at half the rail voltage.
The resistor between the TLE and the opamp is a gate stopper to prevent oscillation. The opamp feedback capacitor is there to shunt high frequencies around the opamp to create unity gain at high frequencies. Again to prevent oscillation. Otherwise the opamp is allowed to use all of its gain to maintain the VG reference as the transistors source and sink current.
Lighting the Epsilon 12
Information about the ε12 delay and offset detector can be found at AMB ε12. The Compact Tube Hybrid adds a lighting circuit that powers two LEDs, one for ON and one for OFF. Here is a partial schematic.
The circuit is simple. The PNP devices are referenced to the V+ rail. When the relay is off the collector of the MPSA14 is pulled up to the rail. This turns off the first BJT pushing its collector to ground. This in turn supplies base current to the second BJT turning it on and lighting its LED. When the relay turns on the collector of the MPSA14 pulls to ground turning on the first BJT and lighting its LED. At the same time the first BJT pulls the base of the second BJT up to the V+ turning it (and its LED) off. The base resistors protect against too much base current and the LED resistors, obviously, set the LED currents. Either two single or one dual LED can be used.
The Power Supply
The CTH power supply is mostly straightforward. The LV supply uses a fixed 24V regulator and the HV supply uses a voltage multiplier identical to the one used in the SOHA II amplifier. The heater supply, however, is different. It uses a switching regulator to drop the voltage from more than 30VDC to either 12.6V or 6.3V. The regulator does this with very high efficiency making it possible for the regulator to run without a heatsink even when supplying 600mA.
Heater Supply - by Chris Forster
Using a 24VAC wallwart to attain a higher 80V B+ coupled with support of 6.3V and 12.6V heated tubes presented an early hurdle for the project. With no center-tap, this meant deriving these low voltages from an approximately 35VDC supply. A transportable amp implies being carried about, and this posed both size and heat constraints. Conventional linear regulation, such as a power resistor followed by a linear regulator, could not be employed.
Our solution is below…. A switch-mode heater supply, which as far as I can tell, could be a first for DIY headamp projects (although not without precedent for DIY audio). Reasons for this may be: 1) common, low-cost switcher implementations (without ripple filters) are less desirable for audiophile applications; and 2) early, lower-frequency devices require larger passive components and could present issues too close to audible range.
My desire to avoid SMD solutions, especially for prototyping, drove us to National's (150kHz) LM2595 step-down (buck) switching regulator. Listening to this implementation in the POC with ripple filter (L1H / C1H), I was unable to detect any audible differences between SM heater and use of a battery for heater supply. The significant ripple reduction we see through use of a post-ripple filter is similar to that seen in figure 17 on page 20 of National's LM2595 datasheet.
National's Simple Switcher tools were used to drive associated component values and selection. The design uses 0.8A parts and supports up to 600ma 6.3V or 12.6V heaters. In feedback, R1H (9k) alone yields 12.6V and switching R3H (7k) in parallel with R1H yields 6.3V. The LM2595 seems a good fit and requires no heatsink in our implementation. Builders will want to follow the BoM components here, e.g. use of low impedance caps, and proper choke and diode values and ratings.
This heater circuit coupled with the use of capacitance multiplier filters in HV and LV sections makes for a small, high-quality power supply.
LV Supply Pre-Regulator Filtering
The only issue with the LV regulator is that a 24VAC walwart, when rectified, makes only about 33VDC. There is not much voltage for a filtering section. In the CTH we've used a 1000μf first filter cap and then a simple darlington capacitance multiplier. This multiplier only drops a few volts but it takes out most of the ripple before the regulator.