1. Home
  2. amps-preamps
  3. 197619-139155

Expanding the Class D Conversation: How Would You Characterize Their Differences?

Expanding the Class D Conversation: How Would You Characterize Their Differences?

I'm currently trialing the NAD M23 (1st gen. Eigentakt-based), and I find it intriguing enough to want to understand it better — which means understanding the broader sonic landscape of class D. So I'm crowd-sourcing.

In a recent exchange, the estimable Ralph Karsten (Atma-Sphere Music Systems) made two comments that stopped me cold. For those who missed it, here's what he said:

"IME, class D amps vary in sound more than tube amps, which is to say, quite a lot."

"IMO there is a bigger difference between various class D amps than you hear between various tube amps. IOW just because you heard one class D amp says nothing about how the next one might sound."

Link: https://forum.audiogon.com/posts/2885828

As I think through this more carefully, these are genuinely important claims. My own experience with tube amps confirms that they produce audibly distinct characters across topologies and designs. If Ralph is right and class D exceeds that range, then generalizing from one class D experience to another is even more hazardous than I assumed.

One specific question for Audiogon members:

If you have a Class D amp or have compared class D amplifiers, how would you describe their character(s)?

Here are some criteria I use:

  1. Frequency balance — Is the tonal response even across bass, mids, and treble, or does it favor certain regions?
  2. High-frequency texture — Are the highs extended and smooth, or edgy, grainy, and fatiguing?
  3. Bass definition — Is the low end tight and articulate, or loose and bloated?
  4. Midrange character — Does the midrange feel present and natural, or recessed and thin?
  5. Transient speed — Does the amp respond quickly to dynamic attacks, or does it sound sluggish and rounded?
  6. Dynamic range — Does it scale convincingly from quiet passages to loud ones, or compress the difference?
  7. Soundstage width and depth — Does it create a convincing three-dimensional image, or sound flat and narrow?
  8. Image specificity — Are instruments and voices placed precisely, or do they blur and wander?
  9. Background noise floor — Is the silence between notes actually silent, or is there grain, haze, or hash?
  10. Long-term listenability — After an extended session, do you want to keep listening, or has something been quietly fatiguing you?

If you can include relevant system context — room, speakers, preamp — please do. Those variables will help me interpret what the amp itself is contributing.

I'm less interested in rankings than in understanding what Ralph mentioned, namely the [vast] range of sonic signatures class D is capable of. Eigentakt, Hypex, Pascal, Purifi, GaN-based, etc. — all fair game.

Price is no constraint here — I'm interested in the full range of what's out there.

hilde45
parkergetdean
1,049 posts
 

@bluethinker I have been using ML for 10 years now, tweaking it for AI, for a client I would not disclose. The AI response is based on the information it is fed, and the algorithms it runs on. You can’t make a gourmet dinner from stuff taken out of the trash. The information on audio gear is wildly subjective, heavily influenced by vendors, sales objectives and pure BS. (On top of the topic being wildly subjective with a funny jargon that gives a normal person a headache) For every honest and expert review, there are 10-20 entertainment piece that will be weighted the same because AI can’t distinguish between relevant and fake (nor its custodians want it to). If it could and would, we would not be fed with lies and propaganda every day in the name of news and facts.

In short, for audio, the sample is so small, AI can just use proper English to sound convincing ("coherent explanation") but it is largely clueless. And it is true for any niche subject. I asked fun questions like best pastry shops in Budapest or best swimming pools in Munich, which movies is more romantic: A or B, the responses were disasters. And THERE IS sufficient info on those like decades and centuries of literature. A couple hours of reading, filtering, iterations and critical thinking yields better results.

It can improve, A LOT, will it, I guess?

yyzsantabarbara
5,557 posts
 

Audio related AI data is not sourced from too many places. Audiogon would be a primary source along with other similar audio related online sources. There are not that many online audio sites the AI LLM models can scrape off.

For example, ChatGPT totally messed up on insisting that the Schiit Yggy DACs can be configured to work on a GeerFab D.BOB. I am assuming some Yahoo online must have posted that it worked. I tried to get it to work and spoke with the designer, and he said it would not work. That was not AI hallucination, just limited data.

AI is based on statistical probability and neural networks. Not sure I call that intelligence.

************************

Tomorrow, I will move the best Class D amp I have owned, the Class D GanFet 6.5 to duke it out with my CODA #16 powering my Yamaha NS5000 speakers. I think the CODA #16 will be my winner, but you never know. I will use both a RME DAC and the imersiv D-1 DAC direct to amp.

I have owned the following Class D amps:

  • LSA Voyager 350 GAN (maybe modded by Ric Schultz)
  • PeachTree GAN1 (modded by Ric Schultz)
  • PeachTree GAN400
  • D-SONIC M3a-800S
  • NAD M22 V2 

The 2 Ric Shultz modded amps were rather good but had weak bass. The 6.5 does not suffer from that with the Magnepan Mini speaker that likes power.

On my office Magnepan setup the CODA #16 is better. It has a bigger stage (especially depth), hits really hard, and of course the Class A smoothness that I love. The 6.5 has that smoothness but not as warm. The 6.5 hits hard but the #16 was sometimes overwhelmingly powerful in my small space. When my imersiv D-1 DAC was fully borken-in the bass on the Magnepan Mini (no subs) was too much with the CODA #16. It is a beast of an amp.

The 6.5 is not going anywhere. I think it is great in the office system, and I will hear how it is in the Livingroom tomorrow.

 

 

 

 

hilde45's avatar
hilde45 OP
6,005 posts
 

@yyzsantabarbara    — Your comparison between the Coda and the GaNFET amp is interesting. Your leaning toward the Coda may actually be a determining factor, no? One question worth putting to you directly: given all the Class D amps you've owned and heard, do you think they are genuinely  very different  from one another, or does a family resemblance persist across them?
  @bluethinker    — My observations on the AGD Audion MKiii and the other amps are coming — I just want to avoid first-thought bias before writing anything up. On the "settling in" question: I'm genuinely uncertain whether that phenomenon belongs to the gear or to us. Changes in what I attend to, listen for, the music I select (which I try to keep consistent), time of day, mood — these are all enormous variables. My hunch is that most listeners overweight changes on the physical side and underweight the psychological variation on our side of the equation. We tend to assume we're the stable, non-changing factor in the experiment — but that's exactly the kind of overconfidence that Harman's careful subjective listening controls were designed to correct for. Just my two cents.

I find the AI conversation here a bit off topic but: 

@parkergetdean    — Whether AI can improve on audio recommendations depends heavily on who is curating and improving the training data. There's also a deeper issue that doesn't get enough attention: the datasets used for audio discussion and review are wildly English-speaking dominant. That means what looks like an "objective" answer is already shaped by a particular linguistic and cultural slice of the listening world. For those curious about this bias in LLMs more broadly, there's some good recent work out of Stanford on exactly this problem: https://news.stanford.edu/stories/2025/05/digital-divide-ai-llms-exclusion-non-english-speakers-research

 

deep_333
2,436 posts
 

Most Class D: Sterility, brittle, thin sounding/lack of body, fatigue inducing

Some GaN class D: sterility is tamed, less fatiguing, lack of body persists 

parkergetdean
1,049 posts
 

re; curating and improving - You can tweak how AI uses the data (tell it that erik_squares is smart, parkergetdean is clueless, Erin is honest, cheapaudioman is shady, etc. ) but until AI can learn to go into hifi shops and listen, I don't know if there is any way to improve the data. That's not to say it can't be useful for beginners, it can work faster in collecting and organizing the data than humans would.

More to discover