I highlighted the importance of the choice of the transistor for bipolar discrete saturation logic before. The relevant parameters were already well understood in the 1960ies and devices like the 2N709 were devised, based on the technology that was available back then.
What about 2020, sixty years later? Clearly, discrete logic is not within the focus of any transistor manufacturer anymore, so we have to pick what is available. What are important criteria (to me)?
Speed - low propagation delay.
Availability - I don't want to work with NOS bought from some obscure vendor. Availability at PCB assemblers would be the best case.
Package size - SMD is a must. Packages smaller than SOT23 would be perfect.
The last two criteria can be easily evaluated from datsheets and vendors listings. It turns out that the first criterion, speed, is not so easy to asses and merits actually testing things with real hardware.
Out of curiosity I tried replacing the base resistor in my 5-stage ring oscillator with a red LED. Why would I do that? Well, first of all the footprint of the LED and resistor exactly matched on the PCB. So why not?
There was a bit more rationale behind this, though: The LED will increase switching voltage and should improve noise margin, similar to what I did with "LTL". Since LEDs have a non-negligible internal capacitance, I was also hoping that the LED would act in a similar way as the reach-through capacitors in parallel to the base resistor, as I tried earlier.
What started as a quick experiment turned into quite a curious feat, as you will notice later.
The schematics are shown above. Please ignore the resistor values, I used 470 Ohm collector resistors in all cases. I built two versions: One with PMBT2369 and one with PMBT3904.
PMBT2369 based RINGO5
Results for a voltage sweep of the PMBT2369 based ring-oscillator are shown above in green, in comparison with previous results. Two things are apparent:
Oscillations are only observed at a much higher operating voltage of ~2.2V instead of 1V for the others devices.
The frequency is continuously rising with the voltage, no maximum is visible within the voltage range.
An increased operating voltage is expected since the additional voltage drop over the LED increases the turn on voltage of each inverter. The larger swing of the logic levels also leads to a reduction of propagation delay and hence oscillator frequency.
The absence of a frequency maximum and the higer maximum frequency compared to the normal ring oscillator without reach-through capacitor (orange) suggest that the LED capacitance does indeed act in a similar way as the parallel capacitor.
It seems that both initial assumptions have been confirmed. Does it mean that RTL with LED-instead-of base-resistor is the way to go? Probably not, as there will be serious issues with device matching when multiple gates are connected in parallel due to variation of LED forward voltage and transistor VBE. This is addressed in the LTL gate, however.
PMBT3904 based RINGO5 with LED instead of base resistor
Ok, let's repeat the same experiment with the "standard" switching transistor PMBT3904.
The plot above shows the ring oscillator frequency when sweeping the voltage up to 7V and then down again. There were quite a few surprises. First, the maximum frequency is above 50MHz, far above anything I had measured with the PMBT3904 before. Then, and this is even more suprising, above 5 V the frequency suddely drops to a very low value below 0.5 MHz - 100 times lower. When the voltage is lowered to below ~3.5 V, the frequency increases again.
The low frequency region is shown in more detail in the figure above.
It is clear that the frequency shows a hysteretic behavior in respect to the voltage sweep. This means that the circuit has two stable operating points between 3.5 V and 5 V - one at low oscillating frequency and one at high frequency. Some waveforms details are shown below.
Waveform of the high frequency oscalliating mode.
Waveform of the low frequency oscillating mode. You can clearly see fully saturated switching.
Waveform in the high frequency operating point close to the switching point. One can see that several frequencies are superimposed and the oscilation gets instable.
This result is quite weird. In one mode, the oscillator is much faster than any PMBT3904 based oscillator i built before, in the other mode it is incredibly slow. What is the exact mechanism behind the metastability? Can it be used for something?
I set up a Spice model in LTSpice to understand better what is going on. See below for the schematics. One big source of inaccuracy is the LED model itself. I took a model for a similar LED from the OSRAM OS homepage and modified it to fit the LED I used (some noname 0603 SMD LED). However, I am certain...
So, we established that the PMBT2369 is the fastest saturation-switching transistor available today and that it achieves a minimum propagation delay of 3.5 ns at a bias current level of 10 mA. How can we go further from there? Obviously, more brute forcing by changing resistor values and increasing currents is not the way to go. Alternatively we can add components to the basic RTL gate.
There are two approaches: Adding a reach-through capacitor in parallel to the base resistor and adding a baker clamp.
The panel above shows circuits of both options individually and in combination including the resulting waveformt of the ring oscillator.
The reach through cap basically shorts the base resistor for high frequency components of the input signal; the rising and falling edge. For the critical case of a falling edge of the input signal it will lead to negative biasing of the base, which will help to remove the saturation charge. You can see above, that it even leads to a negative spike of the output signal due to coupling to the collector.
The baker clamp prevents saturation of the transistor by shorting the base with the collector if the base potential is too high. The choice of the diode is quite critical here: A schottky diode is needed to minimize forward voltage and the capacitance needs to be lower than CBC of the transistor to avoid excessive miller capacitance that could limit the turn-on time. Since the baker clamp shorts the base to the collector, also the low voltage level is increased to about 0.4V. Therefore, noise margin is drastically reduced for this solution, something that is not preferrable considering the already low noise margin of RTL.
Spice simulation of the ring oscillator at a supply of 5V shows that both options improve switching speed and the combination of both measures is cumulative. All options use a 180 Ohm base resistor.
The reach through cap works best, when the voltage drop across the base resistor is high, because this leads to a stronger negative overdrive of the base. This condition is optimizated when Rb is large compared to Rc, as is evident for the strong increase in frequency for lower collector resistance in the plot above.
Due to lack of suitable schottky diodes I was only able to test the reach-through cap in real hardware.
Comparing an RTL oscillator with and without reach-through cap reveals quite dramatic differences. The yellow and blue graphs above are from an oscillator with the same Rb and Rc. The blue one has a 68 pF reach-through cap added in parallel. Due to negative base overdrive, no local maximum of switching speed is observed anymore and the oscillator frequency continues to rise for higher supply voltage/current. A scope trace shows that, despite the high frequency, there is still fully saturated switching. The signals looks a bit smoothed, which may be caused by bandwidth limitations of my probe and the scope (100 MHz):
In parallel to building the ring oscillator models I also implemented the same in LTspice. You can find the files files on the project page. Please find a very brief summary below.
I implemented a 5-stage ring oscillator with an output buffer. Same circuit as on the PCB.. Using the transient function, it is possible to investigate frequency and waveform.
To test the transfer function, I used the DC sweep function in the circuit below. Since the loading of the output has a significant influence on the output voltages of an RTL gate I looked at both loaded and unloaded output.
Sweep output shown below.
An inverter circuit is shown below. As you note, I also did some experiments with a reach through capacitor.
Resistor variation on the PMBT2369 based inverter
I used the ring oscillator model to investigate the impact of base and collector resistor on ring-oscillator frequency and hence propagation delay. the graph below shows the collector resistance on the x-axis and various plots as a function of base resistance.
The 2N3904 is one of the most widely spread switching transistors. There are countless SMT variants available from at least 15+ different vendors in all kinds of packages. Mouser lists more than 50 variants, LCSC more than 80. I picked four devices in SOT23 package at random, trying to balance between more established manufacturers and lesser known chinese sources. Ironically the Nexperia part was the cheapest. The table above lists the four devices and their respective datasheet values. Siko admits to having a slightly slower device while the other datasheets look like they were copied from each other.
Digital transistors look extremely appealing to build discrete logic: They come with integrated resistors so the bias emitter in an RTL gate can be omitted. Furthermore, they are available from many manufacturers in basically any package type, down to chip scale packages, at very lost cost.
There is a nice introduction to digital transistors by Rohm here.
The internal circuit of such a device is shown above. A wide variety of value for R1 and R2 are available. Unfortunately, the minum value for R1 is 1 kOhm, which will already limits switching currents significantly.
Furthermore, although many different variants are listed, not all of those are available. I picked four devices at random from different suppliers.
The candidates are listed above. What is alarming about these devices is that basically no information about transient behavior and switching applications is listed. A few devices liste transistion frequencies, which are rather unimpressive.
As outlined earlier there are still two fast switching transistor types on the market: The PMBT/MMBT2369 and the BVS52. Both are offered by Nexperia and On Semi.
The plot above shows a comparison of both devices. It appears that the PMBT2369 is quite a bit faster. The MMBT2369 shows a maximum at around 5V supply. At higher voltage (current), the saturation charge seems to dominate switching behavior. This trend can only be seen at >8V for the PMBT2369.
Minimum tpd of the MMBT2369 based RTL inverter is 6.8 ns, that of the PMBT2369 based inverter is 4.8 ns.
It's quite difficult to asses the actual origin of the speed difference, just speculating: The higher current gain will increase the miller capacitance of the MMBT2369. However this is mostly relevant while the device is turned on. The critical condition is the storage time of the saturation charge, which flows out of the base when VBE is already zero and the miller capactance plays no role.
The PMBT2369 definitly appears to be the faster of the two. In fact, the PMBT2369 seems to be the fastest bipolar switching transistor that is still available in high volume today.
Why even look at other devices? Unfortunately it seems that the PMBT2369 is only available in the relatively large SOT23 package. To build dense discrete logic (yes, it makes no sense, I know), it would be of interest to use much smaller packages such as SOT523 or SOT723.
Why is it so difficult to identify the fastest transistor for our application from merily looking at the datasheets?
One issue is that many transistor types are not characterized for saturation switching operation anymore and hence little information about that use case is available in the datasheet. As it turns out, many newer transistor types are not optimized for this operational mode either...
A second issue is that most datasheets only list maximum and minimum values. These usually represent garuanteed test limits, but they are not necessarily representative for the typical performance of the device.
In an ideal world, all of this could be solved by simulation. Unfortunately, I had very mixed results using spice models provided by different vendors. In some cases this may be, because the spice model is not finetuned to switching operations.
In the end, the only way is to actually test various candidate devices. The tool of the trade is the ring oscillator.
The images above shows a single RTL inverter including design settings. I used a 2.4 kOhm collector resistor and 360 Ohm base resistor for all investigations. To form an oscillator, an odd number of inverters have to be connected in a circular fasion.
The picture above shows the design I used for most of my investigation. The first five inverters form the ring oscillator, the sixth inverter is used as an output buffer to make sure there is a well defined load on the oscillator.
Btw - the wire was used to fix a flaw in the first revision of the PCB.
The image above shows a smaple measurement from the article i linked above. By changing supply voltage it is possible to investigate the behavior it difference current settings. The propagation deleay of individual gates can be easily deduced from the oscillator frequency by tpd=1/(2*5*fosc).