I have built a prototype implementation of the PFF circuit using the TPA3126 chip. Just to remind you, the preamp stage is tube-based, utilizing nuvistors produced in the former USSR—specifically, 6S51N-V triodes (the "V" suffix indicates a military-grade version rather than a standard civilian type). In my opinion, I have achieved the intended goals. The PCB design will undergo minor adjustments.
During measurements, I used a load consisting of four 33 ohm, 20 W cermet resistors connected in parallel, to which I added a voltage divider composed of two 47 ohm, 5 W resistors and one 27 ohm, 5 W resistor. The measured output voltage was taken across the 27 ohm resistor. This 27-ohm resistor is paralleled with a 33 nF MKT capacitor, forming a first-order filter with a cutoff frequency of approximately 51 kHz.
The measuring device used was the QuantAsylum QA-403.
First, the frequency response. The drive level was set to -3 dB (with 0 dB being the threshold at which distortion begins to occur). The frequency response measurement uses a signal known as expo chirp (it provides a slightly sharper response line in the upper part of the audio band compared to the method based on white noise):
FFT analysis at a frequency of 1 kHz, signal level at -3 dB. Frequency scale shown in both linear and logarithmic formats:
FFT analysis at frequencies of 125 Hz and 8 kHz:
FFT analysis at 1 kHz and a drive level of -3 dB, using A-weighting:
The SNR value was 95.2 dB, and 105.3 dB when using A-weighting. These values were measured with a 1 kHz sine wave input signal at a drive level of -3 dB.
FFT analysis for a multitone test at a drive level of -20 dB:
Time for THD values broken down by harmonic components, at a drive level of -3 dB and for the following frequencies: 125 Hz, 1 kHz, and 8 kHz:
The slight "notch" visible just below 1 kHz in the FFT analysis is, in my opinion, a remnant of a minor instability in the voltage regulation loop of the boost converter that supplies 54 V to the tube-based input stages.
I find it difficult to interpret the results of the multitone test conclusively.
As for the THD levels and the distribution of harmonics, a clear increase can be observed with rising frequency. However, the distortion is lower than the values specified in the TPA3126 datasheet (although the curve needs to be extrapolated as a parabola, since the sharp 24 kHz low-pass filter suppresses the higher-order harmonics).
Once again, I’d like to remind that the input stage is tube-based, which contributes additional distortion — notably, a significant second harmonic component is evident, which aligns with the design goals.
Now I'm waiting for the proverbial cold shower — or perhaps even a bucket of scorn poured over my head. I welcome all comments and feedback.
Best regards, Tomasz.
During measurements, I used a load consisting of four 33 ohm, 20 W cermet resistors connected in parallel, to which I added a voltage divider composed of two 47 ohm, 5 W resistors and one 27 ohm, 5 W resistor. The measured output voltage was taken across the 27 ohm resistor. This 27-ohm resistor is paralleled with a 33 nF MKT capacitor, forming a first-order filter with a cutoff frequency of approximately 51 kHz.
The measuring device used was the QuantAsylum QA-403.
First, the frequency response. The drive level was set to -3 dB (with 0 dB being the threshold at which distortion begins to occur). The frequency response measurement uses a signal known as expo chirp (it provides a slightly sharper response line in the upper part of the audio band compared to the method based on white noise):
FFT analysis at a frequency of 1 kHz, signal level at -3 dB. Frequency scale shown in both linear and logarithmic formats:
FFT analysis at frequencies of 125 Hz and 8 kHz:
FFT analysis at 1 kHz and a drive level of -3 dB, using A-weighting:
The SNR value was 95.2 dB, and 105.3 dB when using A-weighting. These values were measured with a 1 kHz sine wave input signal at a drive level of -3 dB.
FFT analysis for a multitone test at a drive level of -20 dB:
Time for THD values broken down by harmonic components, at a drive level of -3 dB and for the following frequencies: 125 Hz, 1 kHz, and 8 kHz:
The slight "notch" visible just below 1 kHz in the FFT analysis is, in my opinion, a remnant of a minor instability in the voltage regulation loop of the boost converter that supplies 54 V to the tube-based input stages.
I find it difficult to interpret the results of the multitone test conclusively.
As for the THD levels and the distribution of harmonics, a clear increase can be observed with rising frequency. However, the distortion is lower than the values specified in the TPA3126 datasheet (although the curve needs to be extrapolated as a parabola, since the sharp 24 kHz low-pass filter suppresses the higher-order harmonics).
Once again, I’d like to remind that the input stage is tube-based, which contributes additional distortion — notably, a significant second harmonic component is evident, which aligns with the design goals.
Now I'm waiting for the proverbial cold shower — or perhaps even a bucket of scorn poured over my head. I welcome all comments and feedback.
Best regards, Tomasz.
Your distortion spectra do not look that impressive. Maybe your coupling network to the analyzer is the culprit.
As the bridged output is symmetrical, your attenuator should be designed strictly symmetrical.
Otherwise you measure additional THD due to common mode distortion -
which is nullifyed in the bridge configuration.
As the bridged output is symmetrical, your attenuator should be designed strictly symmetrical.
Otherwise you measure additional THD due to common mode distortion -
which is nullifyed in the bridge configuration.
Schema of probe between balanced outputs of power amplifier and balanced inputs of meter - analyser.
The distortion spectrum is a combination of distortions introduced by the Class D power stage with the PFF loop, and those contributed by the tube-based preamp stage.
An additional note: this is not at 1W output power, but rather at around 16W (i.e., -3 dB from maximum output).
The tube stage’s signature is visible in the relatively strong second harmonic, as well as in the third.
The power stage contributes primarily to the 5th and higher-order harmonics.
There is a clear pattern—odd-order harmonics dominate, which I believe result from the influence of the filter inductor cores.
In the simulations, ideal inductors were used.
The higher odd-order harmonics of 1kHz are at around -90 dB, corresponding to approximately 0.003%. For 8kHz are at around -60 dB corresponding to approximately 0.003%.
According to the TPA3126 datasheet, for 16W output power, the THD is 0.02% at 1 kHz and about 0.2% at 6 kHz (contains only 2, 3 harmonic!), based on measurements from the evaluation board.
Furthermore, the higher even-order harmonics are about 12 dB lower, i.e., below 0.0008%.
I would need to modify the test circuit in order to bypass the tube stage entirely.
Alternatively, I might experiment with changing the modulation scheme.
The distortion spectrum is a combination of distortions introduced by the Class D power stage with the PFF loop, and those contributed by the tube-based preamp stage.
An additional note: this is not at 1W output power, but rather at around 16W (i.e., -3 dB from maximum output).
The tube stage’s signature is visible in the relatively strong second harmonic, as well as in the third.
The power stage contributes primarily to the 5th and higher-order harmonics.
There is a clear pattern—odd-order harmonics dominate, which I believe result from the influence of the filter inductor cores.
In the simulations, ideal inductors were used.
The higher odd-order harmonics of 1kHz are at around -90 dB, corresponding to approximately 0.003%. For 8kHz are at around -60 dB corresponding to approximately 0.003%.
According to the TPA3126 datasheet, for 16W output power, the THD is 0.02% at 1 kHz and about 0.2% at 6 kHz (contains only 2, 3 harmonic!), based on measurements from the evaluation board.
Furthermore, the higher even-order harmonics are about 12 dB lower, i.e., below 0.0008%.
I would need to modify the test circuit in order to bypass the tube stage entirely.
Alternatively, I might experiment with changing the modulation scheme.
Last edited:
By omitting the second harmonic (and assuming its value equals the arithmetic average of the remaining even-order harmonics, i.e., -105 dB), we obtain a THD level of 0.005%.
This is four times lower than the value given in the datasheet, which is 0.02%.
The depth of the PFF loop corresponds to a reduction factor of 4.9.
In my opinion, this result shows good agreement with expectations.
This is four times lower than the value given in the datasheet, which is 0.02%.
The depth of the PFF loop corresponds to a reduction factor of 4.9.
In my opinion, this result shows good agreement with expectations.
Yes, because the analyzer—not only containing an ADC for measuring the amplifier's output signal—also includes a DAC that provides the input signal to the amplifier. In this case, the signal is unbalanced (single-ended), meaning the analyzer’s ground is connected to the amplifier.
If the generator (DAC) and the analyzer (ADC) were separate, galvanically isolated devices, then both resistors wouldn’t necessarily have to be identical.
Of course, I used commercial cermet resistors, which don’t have ultra-tight tolerance (e.g., 1%).
Nevertheless, we see that the higher even-order harmonics are approximately 12 dB lower than the higher odd-order harmonics.
For a quick and reasonably simplified assessment, in my opinion, this approach is sufficient.
If I manage to get a QA-451 (load and 7th-order 67 kHz low-pass filter with a balanced measurement probe), I plan to deliberately build a test setup without the tube preamp stage (using OPA862 and TPA3126) and measure it.
In the post https://www.diyaudio.com/community/...timizing-pffb-for-tpa3245.415655/post-8018330, I mistakenly pasted the FFT spectrum for 8 kHz instead of 1 kHz.
Apologies for that.
Here is the correct plot, now with a logarithmic frequency scale.
Apologies for that.
Here is the correct plot, now with a logarithmic frequency scale.
Considering your output level being close to clipping results including tube pre look not bad at all.
Higher harmonics are typical for TPA3118, they will decrease with lower output levels.
Getting lower THD than TI datasheet is to be expected, as your inductors are better than the ones on te EVA board.
And keep in mind that THD spec is a guaranteed limit, not a typical lab measurement value.
Higher harmonics are typical for TPA3118, they will decrease with lower output levels.
Getting lower THD than TI datasheet is to be expected, as your inductors are better than the ones on te EVA board.
And keep in mind that THD spec is a guaranteed limit, not a typical lab measurement value.
I should add that the graphs shown in the datasheet are actually described as EVM measurements.
However, I have a question regarding the multitone test result (performed at –20 dB drive level). I received a comment in response to the graph and would be interested to hear your evaluation of it.
Multitone Test Summary
A 51-tone multitone signal was used to evaluate the linearity and distortion performance of the amplifier under test, which includes a vacuum tube input stage and a TPA3126 Class D output stage with PFFB (Post-Filter Feedback). The drive level was set to –20 dB relative to full-scale output.
- Fundamental peaks: ~–25 dBFS
- Noise floor:
- At low frequencies (~100 Hz): ranges from –92 dBFS to –120 dBFS
- Above 2 kHz: significantly lower floor, reaching –145 dBFS
- Spurious tones: Present between harmonics, typically no higher than –85 dBFS
- FFT window used: Hann (fixed, not user-selectable)
- SNR at –3 dBFS drive:
- Unweighted: 95.2 dB
- A-weighted: 105.3 dB
The noise floor shape above 2 kHz closely resembles the amplifier’s intrinsic noise spectrum, indicating a clean signal path. The slight rise of the noise floor at low frequencies is consistent with 1/f noise and the influence of the vacuum tube stage. The presence of low-level intermodulation products between tones is acceptable and expected given the windowing limitations and test configuration.
Overall, the multitone test confirms very low distortion, high dynamic range, and a clean spectral response, especially in the mid and high-frequency ranges. The performance aligns well with expectations for a properly implemented PFFB topology and high-SNR design.
That is a lot of data - it would be helpful to see the data with tube input stage bypassed and using a reference output level closer to listening volumes rather than -3dB away from max (clip). Say the usual 2.83vrms into 8ohm load (1W). I see the 2nd harmonic at -30dB or -40dB and that is quite high (1%) and I wonder if that is dominated by the tube. I generally like to see THD about 0.06% max. The decreasing higher orders is good though.
Hi y'all, i have to second xrk - these are not very impressive thd numbers.
What i suspect: second harmonic is coming from the tube pre. Tubes are said the produce a significant amount of even order thd, that is what makes their sound.
The third order harmonic can stem from the actual measurement system, especially the load resistors are sometimes tricky due to their voltage/resistance nonlinearity.
Last time i did amp measurements i tried three different load resistors with three different results. The worst ones were the non-inductive ceramic types, some cheap chinese ones did actually perform quite well. (Green wirewound types with 25W)
In the end i did settle for 6 150W types https://www.widap.com/en/products/wire-wound-resistors-wd/
The Multitone Test looks quite ok on the other hand. Did you do dualtone tests?
What i suspect: second harmonic is coming from the tube pre. Tubes are said the produce a significant amount of even order thd, that is what makes their sound.
The third order harmonic can stem from the actual measurement system, especially the load resistors are sometimes tricky due to their voltage/resistance nonlinearity.
Last time i did amp measurements i tried three different load resistors with three different results. The worst ones were the non-inductive ceramic types, some cheap chinese ones did actually perform quite well. (Green wirewound types with 25W)
In the end i did settle for 6 150W types https://www.widap.com/en/products/wire-wound-resistors-wd/
The Multitone Test looks quite ok on the other hand. Did you do dualtone tests?
I had a different experience with the green wire wound resistors - big 100w ones were terrible and added significant THD to measurement. Best resistor is a flat thin film with heatsink capability made by EBC. These are 300W and can be found in 10ohm values on eBay for a reasonable price. They are absolutely the best - I get the lowest THD with them and coupled with a CPU cooler heatsink fins with fan, lets you drive them hard.
EBC UXP-300 is the model I think. They look like this:
EBC UXP-300 is the model I think. They look like this:
The test board is not exactly modification-friendly. On top of that, there are issues with MKS2 Wima capacitors (zero stock availability and even restrictions on sales to Poland, where I live). As for testing, unfortunately the QA-403 software does not support typical dual-tone IMD tests with freely selectable frequencies. Only the multitone test is available. Yes, it is possible to generate signals externally and feed them into the QA-403, but that adds complexity.
There is also one more thing to investigate: the influence of signals outside the DAC's sampling frequency range, as they can potentially introduce spurious tones in the FFT spectrum within the audio band. So, one would need to ground the inputs of the OPA862 and capture the FFT spectrum both with and without power supplied to check for any residual switching artifacts affecting the DAC.
I understand that you are interested in results without the tube preamp stage, so I’ll need to prepare modifications to the test board accordingly.
Regarding the distortion level introduced by the tube: around 1%, mostly second harmonic, with the third harmonic at around 0.1%, followed by a rapid drop-off to levels below 0.01% — in my opinion, this adds a desirable “coloration” that is not perceived negatively by our auditory system. The second harmonic level drops roughly linearly with the decrease in signal level. The output power at which odd-order harmonics begin to appear is around 33W into 8 ohms. The proposed 1W measurement would correspond to roughly 1/33 of that power, or about 17% of the voltage drive. I estimate that the second harmonic level — and thus the resulting THD — would be around 0.2%.
If I may add a personal opinion: very low levels of second harmonic distortion hardly affect the sound character by adding a “tube signature.”
In the meantime, I’ve been working on a revised amplifier concept using a different tube – the 6N28B-V, a frame-grid subminiature dual triode with characteristics similar to rod tubes. This change allows a new topology: one gain stage, a split-load phase inverter, and op-amp-based buffers. Due to the relatively low B+ supply (around 50–55 V), tube-based cathode followers would not provide low enough output impedance for proper PFF symmetry when driving the TPA3126.
The updated design includes a minor frequency response correction in the preamp stage and slight tuning of the PFF loop components. Since loudspeakers are complex, non-standardized loads (composed of various R, L, and C elements), I simulated the frequency response over a wide range of RL combinations as a practical compromise.
The simulation used R values from 3.9 Ω to 68 Ω (E12 series) and inductances of: 1 nH, 1 µH, 2.2 µH, 4.7 µH, 6.8 µH, and from 10 µH up to 6.8 mH (also E12 series). This results in a wide tolerance band for frequency response. The worst-case -0.5 dB bandwidth occurs with L ≈ 30 µH and low series resistance, falling to ~16 kHz. For higher inductance values, the -0.5 dB bandwidth may improve, but a gradual gain drop can already be seen from 1 kHz upward. Things look better when considering the -3 dB point, which gives a more relaxed margin. See the attached graph and simplified schematic used for the simulation.
I’m fully aware that modeling the load as a simple RL network does not fully reflect the behavior of a real loudspeaker system. However, generating a meaningful tolerance field for frequency response based on actual loudspeaker realizations — even for a hypothetical "average" batch of typical two-way or three-way crossover designs — is not feasible within a reasonable time and effort. After all, where would one even obtain a comprehensive dataset covering the construction variants and crossover topologies used in commercial loudspeakers?
So, at some point, a compromise has to be made — even if it's not ideal, it’s a pragmatic approach that still provides useful insights into how the amplifier behaves under a range of reactive loads.
There is also one more thing to investigate: the influence of signals outside the DAC's sampling frequency range, as they can potentially introduce spurious tones in the FFT spectrum within the audio band. So, one would need to ground the inputs of the OPA862 and capture the FFT spectrum both with and without power supplied to check for any residual switching artifacts affecting the DAC.
I understand that you are interested in results without the tube preamp stage, so I’ll need to prepare modifications to the test board accordingly.
Regarding the distortion level introduced by the tube: around 1%, mostly second harmonic, with the third harmonic at around 0.1%, followed by a rapid drop-off to levels below 0.01% — in my opinion, this adds a desirable “coloration” that is not perceived negatively by our auditory system. The second harmonic level drops roughly linearly with the decrease in signal level. The output power at which odd-order harmonics begin to appear is around 33W into 8 ohms. The proposed 1W measurement would correspond to roughly 1/33 of that power, or about 17% of the voltage drive. I estimate that the second harmonic level — and thus the resulting THD — would be around 0.2%.
If I may add a personal opinion: very low levels of second harmonic distortion hardly affect the sound character by adding a “tube signature.”
In the meantime, I’ve been working on a revised amplifier concept using a different tube – the 6N28B-V, a frame-grid subminiature dual triode with characteristics similar to rod tubes. This change allows a new topology: one gain stage, a split-load phase inverter, and op-amp-based buffers. Due to the relatively low B+ supply (around 50–55 V), tube-based cathode followers would not provide low enough output impedance for proper PFF symmetry when driving the TPA3126.
The updated design includes a minor frequency response correction in the preamp stage and slight tuning of the PFF loop components. Since loudspeakers are complex, non-standardized loads (composed of various R, L, and C elements), I simulated the frequency response over a wide range of RL combinations as a practical compromise.
The simulation used R values from 3.9 Ω to 68 Ω (E12 series) and inductances of: 1 nH, 1 µH, 2.2 µH, 4.7 µH, 6.8 µH, and from 10 µH up to 6.8 mH (also E12 series). This results in a wide tolerance band for frequency response. The worst-case -0.5 dB bandwidth occurs with L ≈ 30 µH and low series resistance, falling to ~16 kHz. For higher inductance values, the -0.5 dB bandwidth may improve, but a gradual gain drop can already be seen from 1 kHz upward. Things look better when considering the -3 dB point, which gives a more relaxed margin. See the attached graph and simplified schematic used for the simulation.
I’m fully aware that modeling the load as a simple RL network does not fully reflect the behavior of a real loudspeaker system. However, generating a meaningful tolerance field for frequency response based on actual loudspeaker realizations — even for a hypothetical "average" batch of typical two-way or three-way crossover designs — is not feasible within a reasonable time and effort. After all, where would one even obtain a comprehensive dataset covering the construction variants and crossover topologies used in commercial loudspeakers?
So, at some point, a compromise has to be made — even if it's not ideal, it’s a pragmatic approach that still provides useful insights into how the amplifier behaves under a range of reactive loads.
did you do some square wave testing already with that network? I've had some massive stability issues with the TPA3223 using the network of the 3251/55. interestingly enough, the recommended values do not result in a stable amp even per LTspice Simulation (or borderline stable). I settled for something like that in the End.
Since I'm currently facing an issue with the TPA3126 on my test board (one channel remains permanently muted), and the proposed tests and measurements would involve effectively "butchering" the board — rendering it unusable for everyday use even if the fault is eventually fixed — I've decided to proceed with the following steps.
First, I’ll modify the filter between the amplifier’s loaded output and the analyzer. This will be a second-order low-pass filter using inductors, with a cutoff frequency of around 180 kHz, and approximately 40 dB attenuation at the switching frequency.
I will also replace the laboratory power supply (which misbehaves due to inrush current and unstable startup — possibly causing improper reset of the TPA3126) with a proven switching-mode Mean-Well power supply.
Here’s the measurement plan:
Then, I’ll proceed with two tests involving the tube preamp stage:
If I find that the measurement results closely match the simulation, I’ll consider further theoretical optimization of the PFF loop (including Zobel networks) to minimize frequency response variation across all R-L load combinations from my RL load model set. While this doesn’t represent a real-world loudspeaker perfectly, it should provide a reasonable and practical compromise — a starting point for modeling amplifier behavior under complex reactive loads.
OK — time to go shopping.
First, I’ll modify the filter between the amplifier’s loaded output and the analyzer. This will be a second-order low-pass filter using inductors, with a cutoff frequency of around 180 kHz, and approximately 40 dB attenuation at the switching frequency.
I will also replace the laboratory power supply (which misbehaves due to inrush current and unstable startup — possibly causing improper reset of the TPA3126) with a proven switching-mode Mean-Well power supply.
Here’s the measurement plan:
- Evaluate the measurement chain. First, I’ll capture the FFT spectrum with the amplifier power supply turned off (but the supply leads still connected — only the positive terminal will be disconnected, while the negative remains tied to the PSU). This will help assess both the system's inherent noise and EMI susceptibility — the PCB can act as an antenna — while the grounded negative lead may reveal PSU-related noise, including mains leakage or switching artifacts.
- Power applied, OPA862 input grounded. This will isolate the intrinsic noise of the TPA3126 in the feedback loop (including the OPA862, since a phase inverter is essential for asymmetric input signals). Additionally, this test will highlight whether high-frequency switching noise couples into the DAC input path, as out-of-band signals (even above the DAC’s sampling frequency) can still cause FFT artifacts within the Nyquist band due to intermodulation inside the DAC. This is a known phenomenon — in fact, sometimes it’s deliberately exploited in narrowband spectral measurements.
- First-order low-pass input filter (~86 kHz cutoff). This step evaluates the full PFF loop with a preamp-like filter in place — similar to the tube stage in terms of its filtering effect — providing a frequency response baseline to compare with simulation results, as requested.
- Square wave test. With the input filter bypassed, I’ll observe the amplifier’s square wave response directly at the output. This will provide insight into loop stability and transient behavior.
Then, I’ll proceed with two tests involving the tube preamp stage:
- Tube stage input grounded. This will allow me to assess its inherent noise — including power supply ripple and cathode emission noise.
- Signal applied to tube input. Here I’ll perform full measurements: frequency response, FFT spectrum, harmonic content (THD), and multitone tests. (The QA-403 software unfortunately does not support standard dual-tone IMD tests.)
If I find that the measurement results closely match the simulation, I’ll consider further theoretical optimization of the PFF loop (including Zobel networks) to minimize frequency response variation across all R-L load combinations from my RL load model set. While this doesn’t represent a real-world loudspeaker perfectly, it should provide a reasonable and practical compromise — a starting point for modeling amplifier behavior under complex reactive loads.
OK — time to go shopping.
Since you’re using LTSpice, you're inevitably relying on a very simplified model of the TPA3223 chip. The TPA3220 and TPA3221 series can be seen as a sort of beta version of the TPA3244/3245. The TPA3223 is essentially a TPA3220/3221 with an input stage and control interface similar to the TPA3116/3118. However, these don't quite match the performance of the TPA3244/3245, or even the newer TPA3250/3251.
In practice, the updated TPA3116 derivatives — namely the TPA3126/TPA3156 — deliver nearly the same THD performance as the TPA3220/3221 and TPA3223. I initially used LTSpice and the same simplified model of a TPA device. Unfortunately, simulation results for the PFF loop are heavily influenced by the bandwidth characteristics of the TPA itself. Others have also noticed this, and tried to address it by introducing a first-order bandwidth limiter. That helps a bit, but still falls far short of realistically modeling the actual TPA internal structure.
Ever since Texas Instruments released a SPICE model for the TPA3126 with nearly complete functional implementation, I’ve been able to compare my previous LTSpice-based simulations and have found major discrepancies. Unfortunately, that model is encrypted and doesn’t work in LTSpice. It’s theoretically compatible with PSpice, but works best with TINA, the TI-recommended simulation platform.
Sadly, there are no equivalent (not even encrypted) models available for the TPA3244/3245/3250/3251/3255 series — let alone for the TPA3220/3221 or TPA3223. That leaves us with experimental optimization, which is inevitably time-consuming and relatively expensive.
The TPA chips do implement internal filtering, and of higher order than just first-order. Moreover, the analog signal processing within the PWM modulator introduces further bandwidth limitations — likely in accordance with the Nyquist theorem. If the goal is high performance even without a feedback loop (PFF), the best choices are the TPA3245/3251/3255 series. (I'm excluding the TPA3244/3250 due to the bottom-mounted power pad, which forces lower switching frequencies — and that compromises high-frequency performance, something already evident in the TPA6304 datasheet.)
If lower output power is acceptable, then I believe it’s better to use the TPA3126 or TPA3156 (the latter is better suited for 4-ohm loads), because usable SPICE models allow theoretical optimization of the PFF loop. You can simulate square-wave response too. While the model is very good for frequency-phase analysis, it doesn’t perfectly replicate real-world behavior in transient simulations.
If you're not chasing ultra-high efficiency, you can add a 680 to 2200 ohm resistor across the output terminals. This changes the role of the post-filter RC networks — from snubbers to Zobel networks (though that brings its own trade-offs). Is the ringing a critical issue? Only in no-load conditions — and that added resistor helps tame it.
Keep in mind: square waves, due to their harmonic content, sound awful and are practically nonexistent in music. Also, audio signal sources inherently act as low-pass filters — providing natural attenuation. A square wave at the output signifies deep overdrive, and TPA devices’ built-in protections will likely kick in at that point.
Let me also remind you of a spectrogram I posted earlier in the discussion. As frequency increases, the spectral amplitude drops significantly. So yes — a compromise is inevitable. For typical audio signals, ringing isn’t as problematic. And besides, most loudspeakers include crossovers with inductive elements, so ringing will happen anyway.
To those arguing that traditional class A, AB, or B amps are better: not necessarily. Even vintage amps, revered for their sound, have limited bandwidth at full power due to the inherent slowness of their linearly operating power transistors. And we’re back to the same core issue: a limited-bandwidth amplifier, enclosed in a global feedback loop, driving a complex RLC load. Everything looks fine under purely resistive loads — until you realize some speaker setups sound downright bad with certain amps.
Anyway, I got a bit carried away. To return to the main point: simulations without a realistic model of the internal bandwidth and filter behavior of the chip have limited usefulness. Their results can significantly deviate from real-world performance. Still, I totally understand — chasing the rabbit is more about the pursuit than the catch.
Best regards.
P.S. Of course, to experimentally optimize the PFF loop, it must be done using a realistic load. Unfortunately, applying test signals during measurements carries the risk of overloading — or even damaging — tweeters. One option is to measure the loudspeaker impedance using an impedance analyzer (ideally with the drivers still mounted in the enclosure, although this requires longer measurement leads). If the crossover schematic is known (and the ESR/ESL values of its components can be estimated), we can build a load that approximates the real loudspeaker system as closely as possible. However, the result will only represent a single speaker system — and only that one with reasonable accuracy.
Last edited:
To simplify verification of output filter and feedback
I assume that measuring unloaded output and max low inductive loaded output
produces the window that covers the full range of possible frequency response with all real loads.
Then the square wave test reveals the truth.
With TIs suggested PFFB I fiddled - square wave response became better -
better never really good.
So I ended up with my "differentiating" PFFB where the hole network is replaced
be a single capacitor in the ballpark of 100pF.
This approach does not reduce noise or distortion at all,
but damps the LC output tank the best way - without any snubber.
You will find the according thread in this forum.
Another way to linearize frequency response is choosing a higher LC resonant
frequency - personally I do not embark on this way.
I assume that measuring unloaded output and max low inductive loaded output
produces the window that covers the full range of possible frequency response with all real loads.
Then the square wave test reveals the truth.
With TIs suggested PFFB I fiddled - square wave response became better -
better never really good.
So I ended up with my "differentiating" PFFB where the hole network is replaced
be a single capacitor in the ballpark of 100pF.
This approach does not reduce noise or distortion at all,
but damps the LC output tank the best way - without any snubber.
You will find the according thread in this forum.
Another way to linearize frequency response is choosing a higher LC resonant
frequency - personally I do not embark on this way.
Last edited:
The TPA3223 for me is a must due to the 5V Rail needed. I used the 325x Series before successfully, but the current draw on the 12V rail makes it unsuitable for an effiecient battery powered system. (Especially sonce im using 3 or 4 per Board.)
What i found were stable square wave outputs on some parts, and some which started oscillating with the amp bandwith of ~70khz
Overall, the amp modell used in Spice did show this quite well, the TPAs in my opinion are not very hard to modell. In the end it is a Class D amp with a certain gain and bandwith.
What i found were stable square wave outputs on some parts, and some which started oscillating with the amp bandwith of ~70khz
Overall, the amp modell used in Spice did show this quite well, the TPAs in my opinion are not very hard to modell. In the end it is a Class D amp with a certain gain and bandwith.
However, you would need to precisely determine the frequency response through measurements — and importantly, not just up to 100 kHz, but considerably beyond that. Then, this response would need to be embedded into a simplified model of the class-D output stage. Unfortunately, I’ve attempted simplified modeling of the TPA3116 (and also the TPA3126), and the results differed significantly from what I obtained using the official SPICE model of the TPA3126 published by TI. So, it becomes necessary to validate the behavior in a real circuit and then refine the simplified simulation model accordingly. It’s doable, but I see an inconsistency here — on one hand, there's a strong drive toward precision, yet on the other, the actual modeling of the TPA device itself lacks rigor.
Moreover, you’ve stated that current draw from a battery-powered supply is a critical factor for you. Given that, I believe choosing the TPA3223 might not be optimal. A better candidate for battery-powered systems would be the TPA3156, mainly because TI revised the modulation scheme in that chip, improving power efficiency. However, I must emphasize — that comes at the cost of higher THD. The BD modulation scheme, which offers lower THD, is not available in battery-oriented operation modes.
I also assume that PCB size (and the overall device footprint) matters to you. I speculate that this is a portable, possibly integrated speaker system, and if so, you may want to consider transitioning to a filterless class-D output stage. Devices like the TPA322x or TPA324x/325x are problematic in that regard — they require a minimum of 5 μH inductance in the load path at their outputs. There's a TI document discussing the TPA3136 and its use without output LC filters, which might be worth looking into.
You’ve only revealed that your system is battery powered — but to me, that suggests a small portable speaker or integrated audio system. In such a device, there's no need to obsess over extreme performance, like ultra-low THD. When the amplifier is built into the speaker system, operating without a load is practically impossible.
Anyway, time for me to go buy the missing components and proceed with the measurements.
Moreover, you’ve stated that current draw from a battery-powered supply is a critical factor for you. Given that, I believe choosing the TPA3223 might not be optimal. A better candidate for battery-powered systems would be the TPA3156, mainly because TI revised the modulation scheme in that chip, improving power efficiency. However, I must emphasize — that comes at the cost of higher THD. The BD modulation scheme, which offers lower THD, is not available in battery-oriented operation modes.
I also assume that PCB size (and the overall device footprint) matters to you. I speculate that this is a portable, possibly integrated speaker system, and if so, you may want to consider transitioning to a filterless class-D output stage. Devices like the TPA322x or TPA324x/325x are problematic in that regard — they require a minimum of 5 μH inductance in the load path at their outputs. There's a TI document discussing the TPA3136 and its use without output LC filters, which might be worth looking into.
You’ve only revealed that your system is battery powered — but to me, that suggests a small portable speaker or integrated audio system. In such a device, there's no need to obsess over extreme performance, like ultra-low THD. When the amplifier is built into the speaker system, operating without a load is practically impossible.
Anyway, time for me to go buy the missing components and proceed with the measurements.