Distortion Surface: old measurement, new approach

Distortion Surface: old measurement, new approach

The following metric is typically included for amplifier manufacturers’ specifications:

THD+N < 0.05%

The above specification is incomplete, with operating conditions left out.

2Vrms 300 Ω load 1 KHz THD+N < 0.05%

This is a specification that allows replication of the measurement conditions exactly.

It is my long held opinion THD+N at a single operating point is nearly useless for assessing audio sound quality. Please see the following posts for reference:
https://www.superbestaudiofriends.o...-of-6-amps-thd-n-and-thd-specifications.6194/
https://www.superbestaudiofriends.o...thd-and-thd-n-specifications-and-graphs.5074/
https://www.superbestaudiofriends.o...distortion-vs-amplitude-a-visual-guide.12279/

However, sweeps of distortion over operating level and over frequency might provide better insight into differentiating amplifier distortion behavior.



This is the very first measurement visualization of this type
Using an iFi Audio iDSD Diablo in ECO mode with a 300 Ω load. In the above measurement the red dot represents THD+N at 1 KHz, 2Vrms which measures 0.0016% or -95.9 dB. Not only is it above a reasonable operating level for most listeners and headphones, it is a ridiculous over simplification about the amplifier distortion performance.

An enormous increase in information is presented by the surface which sweeps THD+N over Level dBu and Freq Hz. Observe increase in distortion as level drops. Typical operating levels are -20 to -10 dBu which for many headphones will produce listening levels of 80 to 90 dB SPL.

Consider a very well regarded amplifier, ECP Audio DSHA-3F (original)


With the same 300 Ω load this measurement reveals a drastically different distortion surface. Let that sink in.

Be mindful of scale changes for the graphs, especially in the Distortion dB axis, for best visualizations of nuances.

Rotating the view for nuanced view of the distortion surface


This is only the beginning. Remember THD+N is also an information reduction. Decomposing THD+N into the following:
D2 (second harmonic)
D3 (third harmonic)
4+HD+N (everything else)
will provide additional insight into the amplifier distortion performance.




D2 and D3 distortion surfaces vary considerably from THD+N and from each other. You would have never known without such measurement detail.



Future research might decompose 4+HD+N into 4+HD and N as separate surfaces which would increase the level of information. Many amplifiers are harmonic distortion dominated at higher operating levels while noise dominated at lower levels. Knowing the level of crossover and how much of each are likely to provide further insights.

 

Even the sweeps prior to creation of the surfaces contain much useful information.

The following sweeps are obtained from a DSHA-3F with 300 Ω load

Legend
Yel: THD+N
Grn: D2
Brn: D3
Red: 4+HD+N

50 Hz sweep


100 Hz sweep


500 Hz sweep


1000 Hz sweep


5000 Hz sweep


Composite animation:


Observe the changes in the distortion profiles over frequency.

To recap, numeric specifications at single amplifier operating points are next to useless. Hopefully the above graphics may be appreciated to impart far more insight into amplifier distortion performance.

I will generate more of these graphics as I revisit amplifiers in my lab inventory but they will be slow in coming as the measurements take several hours, the data is exported and manually compiled into a graphing program for surfaces. Screenshots are captured manually and also incorporated into a gif generator.

 

I am too stupid to be able to fully digest this all in a glance but please know I sincerely appreciate your taking time to make this and I'll be reading this thread through a few times to try and wrap my head around the conceptualisation. You are a very cool person. Thanks Bob!

Addendum: I vaguely recall perhaps you sharing something similar in a previous thread, only that the surfaces were colour coded and superimposed. Presuming that'd be too visually congested to be useful since even individual ones are a bit messy, but it seems like that might be SOMEWHAT serviceable with being able to see what's most prominent at a particular SPL and frequency?

 

So the operative question is... do we want our surface distortion plots to look more like stealth fighter jets, or less?...

 

The measurements themselves take hours?

Fair number of programmers on here might be able to eliminate the manual creation of the graphs...

 

I will be attending this AES presentation to congratulate you myself.

 

I leave to the collective imaginations this graphic on the topic:


Created while listening to Art of Vision by Tangerine Dream on DSHA-3F with KTE Spring 3 and HD800-JAR.

 

Truly amazing stuff!

Some other ideas:

Your D2/D3/D4 plots but with phase in place of amplitude along the z axis.

IMD plot: frequency a vs frequency b vs amplitude of frequency a+b, I guess at a fixed amplitude of frequency a and b. The slice along x=y of this is the slice along fixed y of your D2 plot.

 

The sweeps are stepped from +10 dBu to -30 dBu in 2 dB increments. Each data point in a sweep is obtained from a 32K FFT averaged over eight time-domain buffers. Multiple sweeps are acquired from 50 Hz to 5000 Hz. Data from each trace on each sweep must be manually exported at this time. There is a possibility for automation using VB script which is part of the dScope programming language.

 

That was my first thought as well!

 

This is very cool! Suggestion/request for display though (some people are never satisfied): a spectrogram style display, with dBu & frequency on the axes and dB distortion toward the viewer (in the colours), might better enable visualizing differences than looking at 3D surfaces displayed in 2D space. For this sort of display, the more closely-spaced frequency runs the better; but If the data export can be automated and if your graphing package can do it, I'd hope it wouldn't add too much more work for you?

 

I was wondering who would be the first to nail it.

From Kung Fu Panda: "There is now a level 0"

 

I can provide you with the source data for alternate visualizations, but it would need to be a zipfile of .csv files. I don't think I can post that here so I'd need to do a dropbox with a DM link.

 

I think individual harmonics vs frequency as an animation stepped by different levels also works great, but amplifiers are likely much more annoying to measure than speakers in this case. I simply did 70dB to 95dB in 5dB increments for speakers, but a broader range would be needed for amplifiers.

However this way you can see that some distortion products don't change with level and some do. For example 3rd order of the HEDD Type 07 woofer was almost constant with level, dominating the measurements at low SPLs.

See here:
https://www.superbestaudiofriends.o...e-07-an-un-fair-comparison.10708/#post-341220

I don't like to see this kind of behavior and I think it limits low level resolution, maybe something similar can be found for amplifiers.

Also different profiles, like distortion lowering with frequency like we see with the DSHA 3F are likely benign, while the opposite probably wouldn't sound so nice.

Maybe I can rig something up for the original Vali. I'd be interested if it follows a similar profile, though likely not.

 

Nice work on the surface plots! Indeed, lots of work to do those sweeps.

Intuitively it feels there is an upper lattice that bounds what may sound like ass, and that could be interesting to try to empirically derive. In your “enlightenment” plot, for example, I imagine rather than the 0 point, there’s probable a continuum earlier (and not necessarily decaying).

 

This feels so right! Also with the suspicion that some manufacturers may be optimizing for 1 kHz just because that’ll generate glowing reviews.

 

What is missing from this new distortion visualization is the amount of Negative FeedBack (NFB) required to achieve the distortion surface. Consider two Perfect amps having identical, perfect distortion surfaces in all aspects:





Perfect Amp T requires 100 dB of global feedback to achieve the distortion surface while
Perfect Amp S utilizes ZERO NFB.

My hypothesis is Amp S will sound far better than Amp T.