January 2005

The T-Rex SET Amplifier: A DIY Project
Article By Dick Olsher And John Broskie
Click here to e-mail reviewer

  Although long dead, the Tyrannosaurus Rex still roams our imagination and popular culture, savagely preying on lesser beasts and slow men on the big screen, or painted purple and dancing with children on the small screen. T-Rex the amplifier, on the other hand, is neither ferocious nor carnivorous nor purple. Rather, this appellation denotes Toroidal King, and is Dick Olsher's (DO) way of paying homage to Plitron), whose toroidal transformers and chokes dominate the chassis and form an essential part of T-Rex the amplifier.

Some three years ago, the T-Rex was born as an 8-wpc reference-caliber 300B-based SET stereo power amplifier, an amplifier that incorporates the design priorities deemed important by DO. What were those design priorities? Many of the commercial 300B amplifiers that had crossed DO's path either lacked power supply refinement; or cried out for a input and driver stages equal to the majesty of the WE 300B; or weren't equipped with small-signal tubes favored by DO. As amplifiers are little more than a controlled means of coupling a power supply to a loudspeaker, power supply sophistication was deemed a major design priority. The single ended amplifier may be almost one hundred years old, but its power supply does not have to be primitive or meager: Retro for the sake of being retro was never a goal. So too, the signal path leading to the output tube had to meet today's higher standard of clarity, refinement, and subtlety. In addition to the requisite 300B, the tube complement includes four 5687s, a 9-pin dual-triode, (two per channel), one 5AR4 rectifier, and an innovative KT-88/6922 based shunt regulator. However, design complexity was reined in to ease construction. Einstein's dictum that things should be made as simple as possible, but not any simpler, was a guiding principle.

To establish a critical mass, two well-known tube gurus were recruited to help out. Steve Bench (SB) and (John Broski (JB) are widely regarded as original thinkers and have contributed significantly to the DIY community. For the T-Rex project, SB contributed some of the early power supply design ideas and assisted in the refinement of the signal path circuitry. It was JB who suggested the virtual cathode follower output stage and he is responsible for the final power supply design.

A working prototype was built by Ron Cox (RC), which then served as a test bed for various design ideas. Several input stage and driver circuit options were evaluated by RC and DO via listening tests before arriving at a final decision. Three output transformers were auditioned: the Plitron SE-3025 and SE-3050, and the Lundahl 1664. Of the two Plitron transformers, the SE-3025 gave the better impedance match with the T-Rex output stage and was selected for a head-to-head shootout against the Lundahl 1664. Both sets of transformers were mounted on the same chassis, which was wired with a switch and two sets of speaker binding posts that facilitated fairly quick A-B comparisons. The Lundahl was liked for its lush romantic sound, while the Plitron won us over on the basis of its deep bass extension and sonic refinement. For the past year and a half a final T-Rex prototype with Plitron SE-3025 output transformers has been auditioned and tweaked by DO in the context of his reference system. All of the voicing decisions were based entirely on the outstanding TJ Mesh Plate 300B. After much experimentation, DO reports that T-Rex lives up to its moniker in terms of musical passion and drama. "The most musically intense SET amp I've heard to date," reports DO. Not surprisingly, the T-Rex now serves as DO's reference low-power SET amplifier. The T-Rex is now in the public domain for the benefit of the DIY community, however, commercial rights are reserved.

Part of the delay in publishing this article was due to a desire to locate a commercial source for printed circuit boards. Much of the grief and construction problems in any amplifier derive from mistakes made during point-point-wiring. The use of PC boards greatly simplifies a DIYer's task. Believe me (I've been there), trying to debug a "rat's nest" of wiring is as difficult as understanding tax code. There is now a good chance that Audio Oasis in Toronto will be offering at least a partial kit by early 2005.

The Signal Path (JB)

        
Click here for full picture

Where to start? When designing power amplifiers, always work backwards: start with the load, then the output stage, then the driver stage, (then the phase splitter stage, if used), and, finally, end with the input stage. Think of it as cheating, like solving a maze puzzle by starting at the goal and working your way out to the opening.

The T-Rex's output stage at first seems no different from countless other single-ended output stages. Yes, the output transformer and output tube are better than most and the 300B's filament uses a DC power supply, but the real difference is found in how the output stage behaves, as it exhibits lower distortion and lower output impedance than comparable output stages. The difference is found in the added distortion-reducing technique: a plate-to-ground negative-feedback resistor (resistor R15). (See tubecad.com...2001_03.pdf for a more extended explanation.)

This old topological flourish went by the name of "the partial feedback amplifier" and "the inverse feedback pair amplifier." How does it work? Put concisely, the added feedback resistor lowers distortion and output impedance, and extends the frequency bandwidth of the output stage by returning a portion of the inverted signal at the 300B's plate to its grid. Overall, it is an amazingly short negative feedback loop.

Wait a minute, isn't all negative feedback bad? "Bad for which circuit?" is the only honest reply that can be made. If all negative feedback is bad, then we should never use triodes, plate resistors with triodes, or un-bypassed cathode resistors, as each defines a feedback mechanism within an amplifier. Consider this: the hundred OpAmps you hear everyday of your life, directly (from cell phones, walkmans and CD players) or indirectly (through recordings that were mixed, equalized, or amplified by using scores of OpAmps), could not make any recognizable sounds without the heavy use of negative feedback. With most solid-state OpAmps, no feedback would mean no sound.

Negative feedback is like a police force. A nosy, strong, and merciless police force is desirable and necessary when you live in a crime-ridden city; whereas surveillance, 44 magnums, and tough-guy attitudes are unwanted and unneeded when you live in an honest, respectful, and peace-loving hamlet. Tubes just happen to be such good audio citizens (because they hold their own internal feedback mechanism) that we seldom need to hammer and twist their output into line in a heavy-handed manner.

For example, the T-Rex amplifier will function almost as well without the added feedback resistor, with just both a bit higher distortion and output impedance and a bit less extended bandwidth. But then, not that much negative feedback was used in the first place; just a tad. Let's be frank, the problem with most single ended amplifiers is that they are a bit too loose. Of course, like a pair of jeans, being too loose after living with too tight is a welcome relief; but after the initial "ah, that's better" wears off, something a tad snugger is welcome. Or to switch metaphors, having been too super-ego dominated, abandoning oneself to the id is a kick, but eventually, several hangovers and convictions later, something closer to being ego stable is embraced happily. Hence, a small amount of negative feedback is desirable.

Couldn't a global negative feedback loop that extended from loudspeaker connection to the input tube's cathode be used instead? Partially yes, but mostly no. Such a feedback loop would contain many phase shifting elements (such as imperfect coupling and bypass capacitors, Miller-effect capacitances, and the output transformer), which could make the use of feedback dangerous, as too much phase shift converts an amplifier into an oscillator.

At the risk of being accused of being metaphor drunk, a large, all-encompassing, global feedback loop is like communism: on paper, it's the perfect solution; but in practice, the perfect nightmare. The more distinctive, quirky, and independent the elements encompassed, the worse the results. In the single-ended amplifier, the output transformer is the least disciplined element in the mix, being horrendously complex, with parasitic inductances and capacitances and resistances lurking in every corner. Overall, a short negative feedback loop that applies to the fewest elements is best, much as Maoism would have worked best had Chairman Mao kept it to himself, for just the addition of his wife would be too much for any negative feedback system.

However, nothing comes free and the price to be paid for this enhanced performance is that transconductance must be used to fuel the feedback mechanism. In other words, without the feedback resistor, the output tube primarily uses its transconductance to create voltage gain by varying its current conduction into the transformer's primary impedance; with the feedback resistor, more of that transconductance is used to lower both distortion and output impedance at the expense of voltage gain. Fortunately, the missing voltage gain can be had from the driver and input stages; although one stage would be pressed to provide sufficient gain, two cascading stages provide almost too much gain.

In the final design, cathode bias is used since it is safer than fixed bias — the cathode resistor helps prevents the output tube's current running away. And it also mitigates much of the output tube's drift over time. Furthermore, it also makes for a softer clipping characteristic.

The T-Rex input stage is a cascode circuit. "CASCODE" is a contraction of "CASCaded triodes with the gain of a pentode and the low noise of a triODE." Cascode circuits can be made from tubes, transistors, JFETs, MOSFETs, and any mix of device technologies. A purely tube-based cascode circuit requires that one triode be in current series with another triode. The bottom triode's grid is fed the input signal, while its plate is loaded by the top tube's cathode. Loading the bottom triode's plate with a cathode prevents the bottom triode's plate voltage from moving very much in response to signal at its grid, as the top tube's cathode voltage is pretty much fixed by its grid voltage. Nonetheless, the bottom tube will experience variations in its current conduction in the presence of an input signal, just as it would if the triode were connected to a regulated power supply.

In a Cascode circuit, this varying current through the bottom tube must also flow through the top tube, as they are in series. Moreover, as the top tube's plate is loaded with a plate resistor, the varying current defines a varying voltage across its plate resistor, which in turn yields the output voltage and gain of this circuit, which can be considerable, often much greater than the triode's mu, approaching Gain = gmRl, as in a pentode circuit. Thus, "the gain of a pentode." (The secret behind the cascode's greater gain is that it is an elaborate attempt to conserve as much of the triode's transconductance as possible. In a grounded-cathode amplifier, the plate resistor undermines the triode's transconductance, as it allows it plate voltage to vary. In the cascode more transconductance is preserved, thus greater current variation is developed, and thus more gain is realized.)

The circuit, furthermore, functions much like a pentode in that the Miller Effect capacitance is very low, as the grid of the bottom tube is shielded from the inverted amplification at the top tube's plate. It differs, however, from the pentode by the fact that there is no grid current flowing into the top triode's grid as there is with the grid 2 of a pentode and by the fact that the noise is lower than a pentode's. Thus, "the low noise of a triode."

Note that the T-Rex input cascode circuit features a slight twist: the top triode's grid is not grounded, receiving instead a portion of the signal at its plate. This topological change makes the cascode function in something of an ultra-linear mode: less gain than a pure cascode, but less distortion as well. Nonetheless, most of the cascode's low input capacitance and resulting wide bandwidth are retained. The T-Rex's driver stage is a simple SRPP circuit. The cathode resistor is bypassed and the output is taken at the top triode's cathode.

The Power Supply (JB)

     
Click here for full picture.

Power supplies matter. The cleaner, the more unvarying, and the lower in impedance a power supply is the better. What are our choices? Batteries come close, but even the best batteries have their own failings (being easily depleted and bulky, for example). Moreover, batteries are not suitable for most vacuum tube circuits, as 400 volts are hard to get from a reasonable number of dry cells. Although the conventional, passive, plug-into-the-wall power supplies can be bolstered with increased capacitance and chokes, they are far from perfect, as a varying wall voltage results in a varying output voltage and their output impedance is anything but flat across the audio spectrum.

Voltage regulators are both the obvious and usual answer. A voltage regulator is a device that maintains a constant output voltage in spite of variations in either the power supply voltage or the current drawn by the load. A voltage regulator does with complex circuitry what a car battery does by brute force: it maintains a steady and accurate output voltage over a wide range of load current variations.

Regulators fall into two groups: switching and linear. In terms of sales, flexibility, conversion efficiency, and weight to power output, the switching regulator beats the linear regulator. Yet, the switching regulator is rarely used in high-end audio products. Why? Because of noise. Switching regulators operate at supersonic frequencies, 40kHz and much higher; and these high frequencies can easily escape and pollute the rest of the circuit.

Linear regulators, on the other hand, are relatively inefficient; but are also lower in noise, require fewer parts, and are usually much less complex and tweaky than a switching regulator. For example, an excellent 1A, 5-volt linear regulator in a three-pin package that requires few supporting parts can be had for less than a quarter.

Linear voltage regulators further subdivide into series and shunt regulators. The series regulator is the most commonly used regulator — for good reasons. A series regulator is like a "Class B" amplifier: it does not dissipate power unless it required to. For example, if the tubes are pulled from an amplifier, the series high-voltage regulator stops conducting until the tubes are plugged in again. When the amplifier is drawing full current from the power supply, the series regulator is at peak dissipation.

In contrast, the shunt regulator is like a class-A amplifier: it is always on, it dissipates the most power at idle, and it is unburdened when power is shifted into the load. For example, if the tubes are pulled from an amplifier, the shunt high-voltage regulator must increase its conduction (and thus its dissipation) to compensate for the missing current that the tubes had drawn. When the amplifier is drawing full current from the power supply, however, the shunt regulator is at its lowest dissipation.

Each type of regulator finds an optimal load. The shunt regulator works best with load currents that remain mostly fixed, such as single-ended amplifiers and push-pull "Class A" amplifiers. On the other hand, the series regulator is best suited to applications that have a varying current draw, such as "Class AB or B" amplifiers.

As you might have guessed, accountants prefer series regulators and audiophiles prefer shunt regulators, not just because the latter are inefficient, but because of sonic advantages. The series regulator works much like a cascode circuit in that the varying current passing through the amplifier is passed on directly to the power supply. For example, if the amplifier is putting out 1kHz, a 1kHz signal will be superimposed across the power supply's effective series impedance, which can cause problems elsewhere in the circuit. With the series regulator, the B+ voltage remains constant at the regulator's output, but the current through the regulator varies. With a shunt regulator, the B+ output voltage also remains constant, but the current variations in the amplifier are masked completely by the shunt regulator, as the shunt regulator and load draw an almost constant current.

Tube-based shunt regulators offer additional advantages over tube-based series regulators in that they usually find a shorter and cleaner path to the ground, the shunting tube's cathode being close to ground, not floating at the output voltage, which also allows this tube to share the power supply that is used by the rest of the amplifier.

The shunt regulator's principle of operation (in AC terms) is simple enough: the regulator is essentially a single-ended amplifier. The shunt regulator strives to force its output (the B+ connection) to fall in line with the input signal (ground) because of the negative feedback loop that feeds any perturbation at the output back into the regulator's inverting input. In other words, when the noise (feedback signal) on the power supply rail is fed back to the regulator's inverting input and then compared to ground, the current flowing through the shunt regulator will vary in response to the unwanted signal at the power supply connection, and then this current develops a varying and opposing voltage across a resistor or choke that cancels this same signal. In other words, the shunt regulator is effectively a high-gain, high-feedback amplifier with a grounded input and huge (purposely huge) DC offset.

If DC regulation on top of AC regulation is required, the shunt regulator will also need a voltage reference. The DC voltage reference is usually a zener or a high-quality IC voltage reference. Like the DC series regulator, the DC shunt regulator needs three essential elements to work: a device whose current conduction can be readily varied, a DC voltage reference that is AC bypassed to ground, and a control (negative feedback) mechanism to alter the flow of current through the pass device, the shunting device. The shunt regulator below uses only seven parts, but it fulfills the three requirements. Resistor R1 and capacitor C provide the feedback path to the triode, which imparts the varying current conduction; and the zener diode serves as the voltage reference. If the regulator's output climbs too high, the voltage presented to the grid will also climb, which will provoke an increase in current conduction from the triode, which in turn will force the output voltage down. If the output voltage falls, the triode will lower its conduction, which will release some of its downward pull on the series resistor Rs.

If, on the other hand, we are willing to forgo DC voltage regulation and its much greater demands on the regulator, then we can also forgo the need for a DC voltage reference. In other words, an AC-only shunt regulator will still strive to maintain a clean, noise-free output voltage, but not a steady, constant voltage. If the wall voltage sags by 10%, the regulator's output will also sag by 10%. On the other hand, a DC and AC regulator would still provide a fixed output voltage. Wouldn't a fixed, steady output voltage be better? Indeed yes, but at great cost.

Forgoing DC regulator simplifies a regulator's design. In fact, a single tube and few resistors and capacitors are all that is needed to make an AC-only shunt regulator.

In the above circuit, any AC signal present on the regulator's output is relayed to the triode's grid, which causes the triode's current conduction to increase with positive going signals and decrease with negative going signals, which in turn causes a greater or lesser voltage drop across the series resistor. This varying voltage is in anti-phase to the original AC signal at the output, so the original signal is greatly attenuated. How greatly? It depends on the triode used, as the greater the gain developed across the series resistor, the greater the attenuation.

So, are high-mu triodes best? At first it would seem so, but as were are striving for greater and lesser current conduction, not voltage amplification, most high-mu tubes are poorly suited to the task. What really matters is Gm (transconductance), as the greater the Gm, the greater the current swings for the same grid signal voltage. Therefore, the greater the transconductance, the lower the output impedance. For example, one triode might have a mu of 10 and an rp of 1000; and another, a mu of 100 and an rp of 10000. Each triode hasa Gm of 10 mA/v (or as it is often expressed in micromhos, 10,000 µmhos); but finding the latter triode is next to impossible, as most high-mu triodes have high plate resistances, which implies a low gm to match.

The single-triode shunt regulator's disadvantage lies in the limited amount of transconductance that any one triode can exhibit (40,000µS at most; usually no more than 10,000µS). Think of the Gm as being the inverse of the effective shunting impedance of the triode. So, for example, a Gm of 10,000µS equals 100-ohms of shunting impedance, as 1/0.01 = 100. Therefore, using this triode in the previous circuit, the triode would define a 100-ohm resistor in a two-resistor voltage divider and the raw power supply noise would only be attenuated to 1/21th of its original value. The illustration below displays the equivalent circuit in AC terms. Obviously, the greater the transconductance, the greater the noise attenuation.

Much like potato chips, just one tube is seldom enough. Even the highest transconductance triode has a paltry amount of transconductance compared to a solid-state device, such as transistors or MOSFETs. And, once again, it is transconductance — the change in current flow due to a change in grid voltage relative to the cathode voltage — that powers a shunt regulator. Adding a first stage (a voltage amplifier) effectively boosts the gm of the shunting tube by magnifying the error signal presented to its grid. One stipulation is that the extra gain stage cannot invert the phase of the error signal, which leaves out the grounded cathode amplifier, but not the grounded grid amplifier.

The grounded-grid amplifier is a real sleeper of a circuit: it does not invert the input signal phase and offers a wonderfully high-frequency bandwidth, as its grid shields the plate and the ensuing Miller-effect capacitance. The downside to this circuit is its very low input impedance:  (Ra + rp) / (mu + 1). But since the input signal is the noise at the power supply B+ connection, the power supply noise can be easily relayed to the cathode through a large valued electrolytic capacitor. And since the power supply noise frequency is twice the line frequency and since harmonics seldom extend beyond the 5th, exceptional high-frequency transmission is not necessary.

In fact, the usual disadvantage — power supply noise leaking into the output of the grounded grid or cathode amplifier — works in our favor in a shunt regulator circuit and against us in a comparable series regulator. Furthermore, when using a power pentode, such as the KT88, the need to offer the pentode's screen grid a clean AC path to ground also disappears, as triode connecting the pentode works to increase the effectiveness of the pentode as a shunting device.

Still, we have only effectively increased the transconductance of the shunting tube by the gain realized by the grounded-grid amplifier, which is strictly limited by the amplification factor (mu) of the triode, which seldom exceeds 100. How can we get more gain? The obvious answer is to cascade a few triode gain stages. While this approach does yield a lot more gain, it also brings potential instability, as each additional triode adds phase shift. In a high-feedback amplifier, which the shunt regulator effectively is, phase shift is dangerous, for if the feedback signal ever returns in phase, the regulator transforms into an oscillator.

How do we get gain without the concomitant phase shift? A different circuit topology holds the key. The cascode topology offers gain (we are no longer limited to the triode's mu) and low phase shift. Unfortunately, also like a pentode, the cascode amplifier suffers from a poor PSRR figure, but in this application, the failing becomes an asset, as the B+ noise is the signal that we wish to amplify, so adding more of the noise in phase to the output can only help — paradoxically enough.

Please refer to the T-Rex power supply schematic shown above. Both channels share the single shunt regulator, which makes no attempt to regulator the DC voltage, only the AC noise present on the B+. In place of a series resistor, a pair of toroidal inductors is used. The shunting tube is a KT88 and a 6922 dual triode is configured in a modified grounded-grid amplifier.

The original grounded-grid input stage was modified to increase its gain by using a cascode topology. First, a second triode and two resistors to bias the top triode's grid were added. These resistors define a simple voltage divider that provides the cascode's top triode with the correct bias voltage. The second step is to add a current-bleeding resistor to the mix. This 50K resistor diverts 6.4mA of current away from the top triode. Because the top triode now sees much less current, a much larger plate resistor can be used, which will greatly increase the gain from this stage.

The secret to understanding how the gain is greatly increased is found in preserving the bottom triode's transconductance. Transconductance, unlike mu, but very much like rp, is not constant throughout a triode's range of current conduction, in that it falls off with decreasing current. Thus, running this tube with a trickle of current so that a larger valued plate resistor could be used will not necessarily increase the gain, as the reduced transconductance will undermine the role played by the larger plate resistor value.

For example in the circuit above, the bottom triode still sees the full 10mA of current, which ensures that its transconductance remains strong. Furthermore, almost all of the bottom tube's current variation must flow primarily through the top triode and its larger than normal valued plate resistor (50k, rather than 22k), as the current-bleeding resistor sees a pretty much fixed voltage across its leads, thus a pretty much constant current draw.

One danger with this trick is that this stage's effectively lower current draw through the plate resistor (3.6mA) will not be able to charge and discharge the Miller effect capacitance of the triode-connected KT88 as quickly as the full 10mA of current would. However, as the most of the noise will be no higher than the audio-band limit (20kHz), this should not be too much of an issue.

Finally, note that the filaments of the 5687 are referenced to a DC voltage of +70V to maintain the heater-cathode voltage differential for the cascode's upper triode within its 100 V rating.

Added

It was brought to my attention that the T-Rex power supply schematic shows an incorrect hookup for the 5687 filaments. What is shown is a series filament connection for a 12V transformer. The correct parallel filament connection for a 6.3V transformer is given in the attached image file. Please use the above image at the end of the article to this effect. My original schematic was correct, but when it was converted to GIF files for publication something was lost in translation.

Construction Notes

1. DO NOT attempt to build this amplifier unless you are an experienced kit builder. Tube circuits, and in particular, HV power supplies, present a risk of serious and even deadly electrical shock. If in doubt, solicit the help of an experienced electronics technician. The authors assume no responsibility for the performance of the amplifier.

2. All solid-state diode bridges should be chassis-mounted internally for proper heat sinking.

3. R8 is a 10K pot that allows for adjustment of the current through the KT-88 shunt regulator. Both the pot and its associated +60 VDC test point should be accessible through the top of the chassis.

4. Because the 5687 filaments are current hogs, each consuming 0.9 A at 6.3V, a separate filament transformer is necessary. We recommend a Plitron (model 7800, 6.3 VAC/4A) filament transformer for the 4 x 5687 filaments.

5. Both of the power supply chokes (L1 and L2) are Plitron toroidals (You should expect nothing less from T-Rex!).

6. The Plitron PAT-22 chassis is good looking and neatly complements the T-Rex project.

7. The various pinout and base diagrams are shown below.