Thực tế về sự hấp dẫn của Audiophile? Trò chuyện về GPU (dòng nVidia Blackwell RTX 50XX-series), suy nghĩ về khung hình AI (“fake frames”), gaming và nghệ thuật như sự tưởng tượng.
I hope you’re all doing well as we enter the year. It’s rather amazing that the first quarter of the 21st Century has passed so quickly! Time flew. I do hope that as time passes, audiophiles worldwide are gradually becoming more knowledgeable about the science behind this stuff and in general much more critical thinkers, more resilient to the snake oil and scams of yesteryear. I simply think that this is part of the evolution that must happen for any hobby in order to not stagnate. If something is contentious or seems “too good to be true” based on conventional knowledge as we all become versed in how audio reproduction works, the product very well could be something to avoid; instead, better to save money and look for something more meaningful to purchase.
As per my usual routine, every once awhile I’ll publish an E-mail or two from readers of the blog and take time to discuss matters brought up that might not have been fully covered previously, often hitting other short discussion items that don’t need a full-length post.
Let’s have a look at this E-mail from Michael in Sweden:
Friday Dec 20, 2024
Hej Arch,
Great stuff. Really enjoyed reading your GPU article. A recent, accidental or not, (they have done it before) release by Zotac suggests that the 5090 will have GDDR7 memory with 32GB of storage. Time to start saving!
These past weeks I have come across some hi-fi related events that I thought might be of interest. Maybe a topic for a future post?
Steve Guttenberg in his best speakers of 2024 video has rated the SoundArtist LS3/5a as worthy of being on his list. A British Audiophile also praised this clone of the legendary BBC LS3/5. Well this speaker was humiliated by Amir on AudioScienceReview ending his review with the scathing line, “I can’t recommend the SoundArtist LS3/5A. Please spend your money on a proper speaker.” So, two subjective reviews and one objective with all the required measurements to back his verdict. How can they be so diametrically opposed?
Similarly Erin found that the Borresen X3, also favorably reviewed by many, to have shortcomings that do not justify the asking price. Again, Subjective vs. Objective. The discussion with Erin can be read here.
Cheers! Michael from Sweden
Thanks for the E-mail Michael and the suggestion for discussions across the various tech topics. Lots of interesting GPU news this month from CES2025 which I’ll leave later. Let’s start with the audiophile issues first.
Indeed, there is a long history of subjective reviews and objective findings questioning the claims of such reviewers. The earliest one I documented on this blog was back in late 2013 examining the “case study” of the Wadia 121 Decoding Computer (DAC). In that instance, we saw all these reviewers in 2012 claiming that this DAC was amazing, capable of taking on other top DACs back in those days, and how unlikely anyone would be disappointed in the sound, etc. And then by July 2013, Stereophile publishes their measurements and we see that other than AES/EBU input, this DAC has issues like failing to pass along hi-res 24-bit data over the other inputs (including USB), uses a poor anti-imaging filter, and John Iverson described poorer resolution/space compared to the Benchmark DAC in comparative listening. So, even before independent review sites like ASR, clearly objective measurements by Stereophile provided a “sanity check” (or at least high-fidelity playback check) beyond purely subjective opinions.
It’s interesting that Guttenberg says he doesn’t particularly like the original BBC LS3/5A but “loves” these China-made clones! Assuming Amir’s samples (did he only test just 1 speaker?) are representative of these, I guess pronounced bass resonance and >2% harmonic distortion across a very broad 2.5-7kHz at low 76dB SPL “speaks” to Mr. Guttenberg! Supposedly he “heard” huge soundstage, mid-range is “drop-dead gorgeous” – “shockingly good“, “not exaggerating“, “I love it more than any real LS3/5A“, “spectacular“, “the one to get“. Bold claims. Are you sure, Steve?
As for The British Audiophile, unlike Guttenberg, he “feels” the SoundArtist clone sounded like an official LS3/5A based on memory of others he had heard in the past without a sample to directly compare. Hmmm, we have to be careful with audio comparisons based just on longterm memory, eh?
It’s amazing that the SoundArtist LS3/5A had such low measured impedance down at 1.75Ω around 150Hz! BBC LS3/5A’s are supposed to be around 6-8Ω minimum. Again, assuming Amir received a decent unit, The British Audiophile obviously was very wrong here as he has no ability to double check the claims from the manufacturer. Even if the drivers were low-quality and the acoustic measurements sucked, it’s pretty bad form to not get the crossover and impedance at least close to spec given the >$600 price. In any event, those review videos did what they were designed to do: sell a few more speakers, get some likes, grab some subscribers, give the viewers some entertainment value, and achieve some monetization. Everyone’s happy, right? (Maybe not those who actually bought the speakers and realized that this doesn’t sound like a proper BBC LS3/5A, with perhaps better ears/brain than Guttenberg?)
To me, beyond just casual “shooting the breeze” sharing of subjective impressions with a buddy, there is almost no enduring value in purely subjective evaluations of audio gear in more serious publications for high-fidelity gear. My belief is that all reviews, in order to be taken seriously in the 21st Century, must incorporate both subjective and objective evaluation to ensure the product in front of them demonstrates quality engineering, commensurate with asking price, before taking the responsibility to make a suggestion that others purchase such a thing.
By the way, for those interested still in these inexpensive China-made LS3/5A knockoffs, over on AliExpress there are all kinds of options. The SoundArtist is still available for US$630, and there are even less expensive “replicas” like this for US$484 or this Anubis LS3/5A for $470, even cheaper, this one asking less than US$400. For those who don’t mind spending a bit more, the BestVox LS3/5A might perform much better at US$800 with wood enclosure, free shipping, described as being made with a “rigorous reproduction spirit”; they have some measurements to show as well. It’s actually tempting to give one of these a try. 🤔
We’re looking at a 50 year old speaker design this year – the LS3/5A speaker went into production starting in 1975.
While I’ve heard LS3/5A variants at audio shows in the past, the closest I’ve tested is the larger vintage genuine Spendor SA1 a few years ago which I quite enjoyed. It’s a larger closed speaker with 6″ mid/woofer for deeper bass extension.
As for the Børresen X3 Erin’s Audio Corner video, yeah, I saw that one. Alas, I never got to hear those speakers at PAF’23. Well, I guess for a high-end brand that makes 6-figure flagship speakers, these are “only” US$11,000/pair speakers… Can’t expect too much, eh? Wouldn’t it be awesome if Erin could measure their $550k M6? Surely those must measure amazingly! 😬
That Mend It Mark fiasco is really something isn’t it? I see that Louis Rossman has put it up on his channel as a middle finger to Tom Evans Audio trying to suppress sharing of information:
Wow. What a mess of a preamplifier internally. Hilarious attempt by the company to remove the component markings! They’re asking US$30,000 for that?! A nice example of these small boutique brands posing as “high end” audio outfits charging huge sums.
Mark seems like a really decent guy and a class act for being so diplomatic and humorous in his video considering what is so obviouslynot worth the money. And look at all that work he did with reverse-engineering the product schematics. (Reverse-engineering is generally not illegal as an independent means to get information from a product that is legitimately obtained.)
Anyhow, I suspect Tom Evans Audio is cooked after all this negative Internet exposure. I can certainly understand why small “high-end” companies like these don’t want people to look inside the enclosure or find out that they’re using off-the-shelf parts and probably generic circuitry to maintain the aura of their designs being exotic. If a company has the audacity to ask for 5-figure prices, then they better have the courage to handle it when these are exposed for all to see.
While I think Danny is off his rocker when it comes to nonsense stuff like cable risers (and other things he has put up videos on), there’s a lot of good sense in the Wilson-related video. The idea about just putting a good amp into the lower woofer section to make it into a hybrid active bass (Puppy) + passive mid-tweeter (WATT = Wilson Audio Tiny Tot) speaker sounds good. The 2-piece design already is screaming for this to be done. Who knows, maybe back in the day when David Wilson was designing these, he had ideas already but didn’t think he could find an appropriate amp to stick in there. A high quality, high-powered modern Class D amp (no need to be too expensive of course!) to power those bass drivers should sound great.
Note that the W/P System 8 is an older model and you can see the Stereophile review from 2007. It was in production from 2006 to 2011 until the recent resurrection in 2024 at $38.5k/pair. Hopefully modern Wilson Audio speakers measure with smoother frequency response after all these years. However, I see that Wilson speakers tend towards low-impedance designs typically with minimum values down around 2Ω and even slightly below – see measurements for the Sasha DAW (2019), SabrinaX (2021), and Alexx V (2021). The Chronosonic XVX (2021) even reaches below 2Ω across a number of frequencies up to 3kHz! One needs to be vigilant selecting the amplifier to pair with these and be mindful of extremely low EPDR which could get some amps to go into overcurrent protection with certain music and especially at higher output levels even if officially rated for 2Ω!
——————–
Switching gears out of audio for a bit, let’s talk about the GPU stuff. Indeed the nVidia RTX 5090 looks like a beast with 32GB DDR7 VRAM – it better be offering top-of-the-line performance at the US$2000 price point! I managed to watch Jensen Huang’s keynote live at CES2025 and was not surprised about the controversy that erupted afterwards surrounding this slide:
For those who might not have seen this, clearly it’s not possible by pure number-crunching that the RTX 5070 could be the equivalent of the RTX 4090. Huang is simply referencing the ostensibly all-important frames-per-second (fps) number and comparing the RTX 5070 with Deep Learning Super Sampling (DLSS) 4 multi-frame generation (the computer uses machine learning techniques to create up to 3 intermediate frames between individually calculated ones) to the RTX 4090 presumably running DLSS 3 (can create 1 intermediate frame at a time). Who knows if ray-tracing is even turned on with this claim. This is an example of using headline ‘benchmark’ numbers without explaining context to everyone to create an impression with some shock value. I’m sure it’s intentional but the effect would not impress knowledgeable tech reviewers, so hopefully nVidia recognizes that they should not do this in the future because the response is not complementary.
Obviously, it’s important that core framerate needs to be decent enough so that the gameplay is adequately responsive. A game with core framerate of 15fps with 3-frames generation, temporally interpolated up to 60fps, is probably not going to be impressive with first-person shooters even if every frame flows smoothly. I believe it’s best to still target around 60fps core. Having said this, I think DLSS 4 + Reflex 2 “Frame Warp” (decrease latency despite multi-frame generation) could be a nice combination. The true test of gameplay latency will be to hear from the Esports players if Reflex 2 works decently enough for them and at what frame rate thresholds. As a more casual player using wireless controllers, I would not likely be able to appreciate much difference!
Hopefully the AI upscaling (CNN to Transformer models for “Super Resolution” and “Ray Reconstruction”) upgrade is as they showed which will also benefit the previous generations, including my RTX 4070 Super, with better image fidelity and possibly less memory use.
I’m still a bit unclear about the Neural Materials and Neural Texture Compression in the RTX Kit. It’ll be interesting to see the kind of texture detail they can achieve out of lower VRAM cards. Thus far, discussions on the technology like this one look impressive.
As noted in my RTX 4070 Super article, I’m not particularly impressed by the RTX 5070. I don’t think there’s any real upgrade. However, I might get my hands on the RTX 5070 Ti (supposedly US$750 MSRP) for some LLM inferencing in one of the work machines. The 16GB VRAM + FP4 “minifloat” hardware support could result in some nice performance with quantized LLM’s. Anyhow, if I do this, I’ll make sure to borrow it for a few days and see how it performs in my Game box at home!
Hmmm… Are these real frames?
On a more philosophical note, about this multi-frame generation idea for gaming, I don’t think there’s any issue. I know some commenters have called it “fake frames“, the technique being “chicanery“, they’re “lying” and other similarly disparaging comments (let’s not get freaked out until we see some testing!). So long as the game plays well and the graphics look good with minimum artifacts, what does it matter?
In some ways, this is like when hardcore audiophiles complain about MP3/AAC and other lossy audio formats (including EAC3-JOC for Atmos) that reconstruct the PCM audio data in ways very close to the threshold of human hearing as to be indistinguishable at adequate bitrates. Well, we’re seeing a form of “lossy” 3D graphics reconstruction here where the compression isn’t about saving bitrate, but reduction of complex calculations such that each frame isn’t freshly being computed, but rather inferenced, leveraging the power of algorithms that have been taught how graphics are supposed to look. At its heart, this is what it means to put “intelligence” to work in a way that’s much more efficient than raw computation.
This by the way is also how our brains work. Our minds interpolate and “fill in the gaps” all the time! For example, our vision fills in the optic nerve physiological blind spot such that we don’t even notice that we’re always missing a piece of our visual field (perhaps like the ‘inpainting’ that’s used by Reflex 2 to fill up the corner blind spots during motion). Likewise, the visual system has eye-to-brain latency, limiting the actual sensory data available to our mind, and so 30-60fps motion still looks smooth.
It has been estimated that minimum visual latency is around 13ms (77fps) even to the point of decoding conceptual meaning! Extrapolating this, it might not be completely unreasonable to suggest that gaming at a rock solid locked 80fps synced with the monitor’s framerate to prevent judderis all we’ll ever need. This is a bit like in the audio world we generally accept limits like 20Hz-20kHz for audible frequency response, 16-bit resolution generally is more than needed for dynamic range, that small temporal jitter from a DAC is undetectable, etc. (as per the article years ago on the limits of human hearing).
To me, the idea of the computer/GPU “imagining” a few frames, literally milliseconds like the 33ms between 30fps, of the graphical world isn’t a big deal since our own minds imagine much of what is seen and heard anyway as the limited data from our sensors (ears, eyes, skin, proprioceptors, etc.) are fed to our brains. The brain creates its own model of the world. While hi-fi is about reproducing what has been recorded accurately/transparently, the future of interactive media has always been about constructing an acceptable gaming virtual reality for us to participate in and enjoy with aims of graphical photorealism (when intended). To have the machine “hallucinate” the imagery bookended by actually computed frames should not be in any way unreasonable!
Making stuff up, imagining, even “hallucinating” in AI such as LLM’s going off the rails with false information is nothing new either. Illusions, dreams, imagination, even hallucinations are all parts of the human condition as the neural networks of our minds make linkages based on what they have learned through experience reaching beyond reality-based limits. It is probably this facility to extend ideas beyond the logical that also gives us creativity, the birthplace of art, and when driven to extremes, mental illness – the source of our imagined fears, delusional cognitions, and hallucinatory perceptions.
Any wonder then that an AI LLM can already do a pretty good job coming up with purely subjective “reviews” for audio gear? (Also this.) Throw in a few plausible bits and pieces, all subjective audio reviews (human or AI generated) of reasonably high performance gear are likely not much more than descriptions based on projections of the person’s biases and illusions; rather than an accurate reporting of what’s “heard”. 😮 Which again is why we need measurements of the output from these engineered products.
The Ubiquity of Art – A Nightmare?
If machines can be instructed to “dream”, IMO the “need” for human-created art is in decline. Machines based on these artificial neural networks will gradually take over the need to hire an artist for their skills and creativity in producing “art”. Yes, I’m sure there will always be a market for genuine 100% human-created paintings, music and sculptures, but most art will likely be some combination of human plus machine-assisted production. Truly 100% human-created art will likely be advertised as “artisan-created” that will cost more, and take more time to make; just as a “handmade” pair of speakers will come with an extra price margin, production wait time, and likely idiosyncratic imperfections in the workmanship compared to the ones off an assembly line with high quality control.
Those who make their livelihood from art have every right to be concerned given that we’re not even more than 5 years into the availability of generative AI tools for public consumption.
“I make art, therefore AI.”
On New Years day, I got this from Michael:
Wednesday Jan 1, 2025
PS: Happy New Year!
Came across this today which might be of interest to you. If you subscribe to Amazon or Apple you can listen to their immersive playlists supporting Dolby Atmos or Sony360.
That’s great that Mr. McGowan says “I love surround sound!“, “so superior to 2-channel” and wishes all his recordings could be multichannel. I did not know he was so complementary towards multichannel. There’s really no reason why he can’t start doing some multichannel mixes at Octave Recordsnow – there’s no need for any magic wand. Seems like more of a financial decision based on what he thinks he can sell to audiophiles. Of course with multi-track recordings, one could do a remix for multichannel down the road and achieve some very nice sounding results even if not released today.
I think Alex, the person who submitted the question highlighted an important observation about the fact that 2-channel can only ever produce a “phantom” center. The interesting and unfortunate thing though is that even with modern multichannel music mixes, many still do not use the center channel, preferring a quad (2 front, 2 back) base arrangement rather than full 5-channels (3 front, 2 back) of content. I don’t know if much streamed music employ 7.1 channels.
Multichannel/Atmos versions of albums like the Moana (2016) soundtrack or Fleetwood Mac’s Rumours (1977) that use the center channel well are still unfortunately not as common as I’d prefer.
Stuck in the center for you: don’t forget the importance of a good center channel! (Apologies to Stealers Wheel. Paradigm Reference Signature C3 v.3 – 3-way WTMW shown.)
For those curious about 3-channel stereo with discrete center, have a listen to some of the 3-channel Analogue Productions SACDs released in the early 2010’s like Nat King Cole’s Songs from St. Louis Blues or The Very Thought Of You, Oscar Peterson’s We Get Requests. Also consider the 3-channel releases on Living Stereo SACDs like some of the Fritz Reiner recordings (Dvorak: Symphony No.9, Mussorgsky: Pictures At An Exhibition, etc.). These versions with a “true” center image have become my references for these albums.
There is one area where McGowan is incorrect, that “there’s not a whole lot of music out there that’s recorded in multichannel“. Let’s just say there is a lot of multichannel content already; I don’t think I care if it was originally “recorded in multichannel” because ultimately it’s all about the mix that was created (itself an artistic and engineering endeavor as discussed by John Atkinson this month). It’s great to see (and hear!) the continued expansion of music available in multichannel through 2024. There is certainly plenty of multichannel/Atmos content these days for those of us with systems that can extract the material. In many ways, this kind of growth was what many of us could only have dreamed of back in the early 2000’s when multichannel SACD and DVD-A began to be released. I certainly don’t see the need to lament as McGowan did about music listeners using soundbars or small smart-speakers. It’s great that products are available for all kinds of settings including those meant for convenience rather than fidelity! Whether you have a lo-fi system in the kitchen, wireless headphones paired to your smartphone, or a hi-fi dedicated sound room, the content is there for you to choose: whether it’s background music, the 2-channel stereo experience, or to go “immersive“.
Thanks for the link to the “Mixed on Focal” multichannel playlists! Nice sounding stuff that I’m just starting to explore.
And a wonderful 2025 to you, Michael! Cheers…
——————–
To end, speaking of Børresen above, one of the companies under Audio Group Denmark includes Ansuz, a friend sent me this link for this device:
Yes friends, only for the highest tier Audiophiliacus supremus, upper echelon vinyl lover, this is the record stabilizer you need weighing in at 0.8lbs, with 12 “interlayer balls” made of titanium, and costs a mere US$8.3k.
I don’t know what to say… Is this awesome? Of course a review for this thing already exists; “Scrupulous honesty demands I award the Ansuz Acoustics T2S Stabilizer an Audiophilia Star. Price, be damned” even! I’m sure the high-end audiophile world appreciates Mr. Kershaw’s impeccably scrupulous honesty (or is he just feeding neurotic psychopathology?).
As usual, such companies (and reviewers) never show us any objective results to demonstrate that it even makes a difference given the obvious “too good to be true” aura and questionable value. Ansuz, do you mind showing the consumer what kind of reduced resonance this puck is achieving? How about a comparison with the simple Pro-Ject Record Puck PRO (a lowly nickel-plated aluminum lump for an obviously inferior US$69). 🤔
I think it would be fascinating to know how many percent this has been marked up compared to production costs. I suspect a very impressive number that few other high-end manufacturers dare to attempt!
Thinking out loud, 1000 5mm titanium ball bearings can be had on Amazon for US$150. It could be a fun project creating something like this with the ball bearings underneath. Heck, maybe you can get even better resonance control by interleaving those 5mm balls with 3mm titanium balls – $150 for 5000 – in a spiral pattern that complements the rotational centripetal forces as the LP spins instead of that boring hexagonal pattern. The “2x resonance controlled double-ballz energy-flo” model should easily be worth US$11.235k because I think the story sounds pretty good; and some of those rich audiophiles really love a great sci-fi story, right? 😁 (Ansuz, if you use this idea for the Record Stabilizer Supreme Mk II, all I ask is acknowledgement. Thanks.)
