Accurate Sound's Hang Loose Convolver Multichannel on low power Mini PC. The disgraceful AQVOX ethernet switch. And some E1DA upcoming evaluations.

Over the last few weeks, I've been using the inexpensive Beelink Mini S Intel Celeron N5095A PC in my music system. Recently, I posted on using it in Linux as a streamer with HQPlayer upsampling and for multichannel Roon capabilities. As you can see in that post, for audio purposes, even this level of CPU performance on a modern entry-level machine is enough for what would be advanced PCM upsampling (ie. 1M-taps realtime PCM) and even IMO very good PCM-to-DSD256 performance. More evidence to remind us that there's generally no need to get excited about some of the more expensive "audiophile" computers and streaming machines which are often embedded low-power computers at heart.

Recently, my friend Mitch Barnett of AccurateSound.ca wondered if I could look into the performance of his latest version of Hang Loose Convolver (HLC) with this low-power computer using my power limited and quiet fan BIOS settings. HLC is an advanced standalone, VST3 plug-in, or AU (Mac) plug-in that provides the ability for audio lovers to load in various convolution filters, allows for easy level matching, and provides basically immediate switching (0.1ms) between these filters for accurate A/B listening comparisons. Let's take a look at this on the Beelink.

HLC currently is available as Windows or MacOS versions. I know Mitch is looking into compiling it for Linux at some point, maybe even Raspberry Pi, we will see. Therefore, with the dual-boot Beelink Mini S, I started it in Windows 11 for a look.

As I mentioned above, HLC has various operation modes, but for simplicity, I'll pair it with JRiver Media Center as audio player. JRiver versions creep up in number over the years, and since I am primarily a Roon user these days, alas have not updated to the latest version. I'll be showing JRiver 28 in this article. In my discussions with Mitch, he recommended using the latest JRiver 30 as there have been bug fixes over the years; for example, I noticed cosmetic plug-in draw refresh quirks in JRiver 28 which did not affect the function for the plug-in.

First, I download the latest HLC from the website and installed the software. It costs US$129 to buy the full license. E-mail Mitch for a 14 day trial.

Since I'll be using JRiver under Windows 11, let's make sure the VST3 plugin is recognized by JRiver to get it ready for use:

As you can see, by default, the installation puts the plug-in into "C:\Program Files\Common Files\VST3", so click on "HLConvolver.vst3" and it'll get recognized in JRiver.

Now, I would strongly recommend increasing the number of VST buffers. The way HLC works to allow for multiple filtersets to operate and switch in realtime is that it runs each of the convolution configurations you load in concurrently. So it will need more CPU processing power and data flow to the plug-in as you increase the number of loaded "Filtersets". Default is 256 buffers and I recommend increasing this to 2048 as a start. Latency would increase by doing this but it would not be significant if we're just using it for audio playback (latency more of an issue if you need audio/video synchronization when playing a movie; in that case, minimum phase filters should also be used).

Increase VST3 buffers to 2048.

Next, we can take a look at a simple configuration .cfg file; let's start with a straight 2-channel convolution for stereo material. In practice, a 32-bit, 48kHz, 65k-taps filter has more than enough resolution for room correction. HLC is able to scale that to whatever playback sample rate you're using. Since it's good to err on stressing the processing demands to make sure there are no issues, let's start even higher with a 96kHz impulse file, 128k-taps, with 2-channel room correction.

The .cfg file format is quite capable of more complex processing such as crossfeeding, implementing digital crossovers, as well as simpler relative level and temporal shift parameters, based on the Sound Forge syntax. Have a look at AccurateSound's operations guide for HLC for some other .cfg examples:

96000 2 2 0
0 0 0 0 0 0
0 0 0 0 0 0

Archimago_FIR_Filterset-2.0-96k.wav
0
0.0
0.0

Archimago_FIR_Filterset-2.0-96k.wav
1
1.0
1.0

The filter file "Archimago_FIR_Filterset-2.0-96k.wav" is located in the same directory as that .cfg file. Okay, let's load that up and play something in JRiver, listening for glitches and capturing the CPU performance:

As you can see I'm playing a 24/192 demo track from Soundkeeper Recordings (top orange highlight) through the system with HLC plug-in active (right panel) showing the 2-channel output levels, barely needing much CPU processing at 15% through the minute of playback (the increase at the end of the graph jumping to 34% is just the screen grab increasing CPU demand). We can see that the plug-in has been loaded with the higher demand 2-channel 96kHz/131072-taps filter.

This tells us the Beelink Mini S Celeron has no problems running a single instance of the stereo 96kHz/131k-taps convolver playing hi-res 24/192 music. Good start.

Let's load up all the Filtersets with this 24/96, 131k-taps .cfg to the VST3 plug-in now and see if this changed the CPU load:

For some variability, even though I used the same filters, I changed the output volume a little for each filterset so I can confirm audible difference when switching.

Indeed, as you can see, the CPU demand jumped up because 2-channel x 6-filtersets = 12 convolvers are now running concurrently - we can instantaneously switch between the settings in realtime. This is how you can test out variations for example between filters created by Acourate or Audiolense or REW or RePhase (some discussions on the latter 2 here). Filter variants that change parameters such as minimum phase or turn off excess phase correction would also be fun to test out; doing this, you can listen for yourself the effect of time correction.

Total CPU load is now around 40%. It still sounds fine with no "blips" or playback errors unless I'm doing other tasks on the machine.

With recent versions of HLC, the plug-in can now do MULTICHANNEL. So then, in JRiver, let's make sure to set the output to my HDMI receiver:

As you can see, I've selected my HDMI output (the TV is a Visio 75" P75-C1) which goes through the Yamaha multichannel receiver, WASAPI driver.

And I can specify 7.1 channel format in the JRiver output:

And now let's use another simple .cfg pointing to a 32/48 7.1 (8-channel), 65k-taps impulse .WAV file this time:

48000 8 8 0
0 0 0 0 0 0
0 0 0 0 0 0

Archimago_FIR_Filterset-7.1-48k.wav
0
0.0
0.0

Archimago_FIR_Filterset-7.1-48k.wav
1
1.0
1.0

Archimago_FIR_Filterset-7.1-48k.wav
2
2.0
2.0

Archimago_FIR_Filterset-7.1-48k.wav
3
3.0
3.0

Archimago_FIR_Filterset-7.1-48k.wav
4
4.0
4.0

Archimago_FIR_Filterset-7.1-48k.wav
5
5.0
5.0

Archimago_FIR_Filterset-7.1-48k.wav
6
6.0
6.0

Archimago_FIR_Filterset-7.1-48k.wav
7
7.0
7.0

Now, load up one of the Filtersets and let's play some standard 7.1 16/48 lossless FLAC which is how I typically store my multichannel rips these days. As discussed years ago, I'm very much "post-hi-res" and will happily downsample/dither my content without concern keeping only a few "true" high resolution albums in my collection based on unimpeded dynamic range and natural looking FFT.

Even though I'm playing 16-bit multichannel FLAC, internally, HLC always performs calculations in 32-bits floating point:

Confirmed 8-channel, 48kHz, 65536-taps filter.

We can see all 8 channels active with this 7.1 16/48, Yello track from the Point (2020) BluRay rip as FLAC. CPU utilization remains very reasonable at only 15% or less.

Suppose we now want to run a bunch of A/B comparisons by loading all 6 Filtersets with 8-channel configurations. This means 8-channels x 6-filtersets = 48, 48kHz/65k-taps convolvers concurrently. This is what the demand on the CPU looks like:

Still very reasonable. Only around 25% CPU utilization with all 6 filtersets loaded! This tells us that if we're sticking with 48kHz/65k-taps 7.1 filters in HLConvolver, we should have no trouble even with the power-limited, fanspeed-reduced Beelink Mini S.

In the 2-channel test above, notice how I pushed the samplerate and tap-length higher. So let's do the same and push things even further for multichannel, what if we feed the convolvers with 32-bit/96kHz, 131k-taps filters?

Let's load this .cfg:

96000 8 8 0
0 0 0 0 0 0
0 0 0 0 0 0

Archimago_FIR_Filterset-7.1-96k.wav
0
0.0
0.0

Archimago_FIR_Filterset-7.1-96k.wav
1
1.0
1.0

Archimago_FIR_Filterset-7.1-96k.wav
2
2.0
2.0

Archimago_FIR_Filterset-7.1-96k.wav
3
3.0
3.0

Archimago_FIR_Filterset-7.1-96k.wav
4
4.0
4.0

Archimago_FIR_Filterset-7.1-96k.wav
5
5.0
5.0

Archimago_FIR_Filterset-7.1-96k.wav
6
6.0
6.0

Archimago_FIR_Filterset-7.1-96k.wav
7
7.0
7.0

Let's play some 5.1 24/96 music ripped from 2L's Divertimenti Blu-Ray:

As you can see, I only loaded 4 filtersets and the CPU demand is up around 40-45% already across the 60 seconds of playback. If I load 1 or 2 more filtersets, the music playback will start intermittently glitching, indicating that this is as far as one should "safely" go using 7.1 channel, 96kHz/131k-taps filters playing 96kHz multichannel content on the low-power Beelink... Or is it? ;-)

Recall above, I only set the VST3 buffers to 2048? How about if we push that to the maximum of 8192 buffers?

There you go. Increasing the VST3 buffers, now Hang Loose Convolver is capable of processing all six filtersets loaded with 7.1-channel, 96kHz/131k-taps settings without high CPU load to cause playback glitch. That's 48 individual convolvers running in the background in realtime on a Beelink Mini S, with power-limited BIOS settings.

A look at the relative sizes of the Beelink Mini S (middle), with the MeLE Quieter3Q (review) top, and Beelink SER4 AMD 4700U (review) below.

So what does this mean?

Simple. For us audiophiles, a low power computer like the inexpensive Beelink Mini S 11th Generation Celeron N5095 CPU is all we really need at this point in history for doing stuff like very high quality filtering (as per discussion of using 2-channel HQPlayer filters previously) or as in this post, running multiple convolution DSP processes even with unnecessarily high 96kHz/131k-taps settings, across the HDMI multichannel output. Rationally, 48kHz/65k-taps will easily suffice for fantastic sound quality.

While there are variables that affect the CPU load like sample rate, number of channels, and VST3 buffers setting, generally, there's nothing to worry about unless you go hard with pushing these to extremes!

If anyone is wondering, no, we cannot do DSP with DSD data unless it gets converted to some kind of multibit PCM first, so the idea of convolution DSP with native SACD/DSD audio is not going to be without conversion losses.

For comparison with something inexpensive like the Raspberry Pi 4, it's no surprise that even my low-power, low-fan speed BIOS-tweaked Mini S easily surpasses the Raspberry performance. It's difficult to make apples-to-apples comparisons due to a multitude of factors (everything from the x64 vs. ARM CPU architecture, to RAM variant, to thermal solution, to OS, and how benchmarks are compiled).

As a point of comparison though, using stock CPU speeds and a cross platform benchmark like Geekbench 5, we typically see the following kinds of numbers:

  Geekbench 5                 Single-Core            Multi-Core
    Raspberry Pi 4 Model B     200-300                600-800  
    Intel Celeron N5095A       625-675                2150-2225

In use, that looks and feels about right for general computing tasks.

DSP processing is more demanding of 32/64-bit floating-point performance. If we look at benchmarks like LinPack, I think the Celeron N5095A is actually much superior with more than 3x performance compared to the Raspberry Pi 4's Broadcom BCM2711. Again, hard to be precise unless we have equivalent apples-to-apples benchmarks and being mindful of optimizations like use of SIMD (eg. NEON, SSE, AVX) instructions. If anyone has seen other good comparison data, I'd love to know.

Furthermore, note that I'm using JRiver and the "generic" VST3 plugin, applying various VST3 buffer settings depending on need. Likely even better performance can be expected if the HLC functions were integrated into a product and optimized as part of its DSP features.

IMO, if we can already achieve all this with current low-priced Celeron CPUs, the future will only bring us even faster, lower-power generations! Likewise, Raspberry Pi-like SBC machines will evolve with greater capabilities - even a jump from the current Cortex A72 Pi 4 CPU to a newer-generation Cortex A76 would give us something between 50-70% bump in Geekbench at the same frequency. Depending on when the "Pi 5" comes, expecting a +100% CPU speed over the current Pi 4 would not be surprising; we shall see.

As previously discussed, spending a lot of money on "audiophile" computers doesn't make much sense. Even home servers managing massive databases, streaming to multiple devices concurrently, and running say Roon with multiple DSP steps (like level normalization + upsampling + high-tap convolution) should not need multi-thousand-dollar machines IMO.

Needless to say, HLC sounds great with my room filters created with Acourate. I've promised Mitch that I'll start looking at multichannel measurements in my room, hopefully this winter, time permitting. And as usual, as much as we get excited about new hardware, make sure to examine your room and its acoustic properties, with the idea of digital room/speaker correction as an optimization step once you've done your best to sort out other clear deficiencies. You might be impressed just how much difference you can make in the sound quality, often way beyond whatever improvements hardware upgrades might bring or certain (more psychological) tweaks like cables could ever do. ;-)

----------------------

"It is the duty of every man, as far as his ability extends, to detect and expose delusion and error. But nature has not given to everyone a talent for the purpose; and among those to whom such a talent is given, there is often a want of disposition or of courage to do it."

-- Thomas Paine, Essay On Dream, 1807

From useful software like HLC, let's talk about some ridiculous scam hardware. As per this video from Linus Tech Tips (I believe these guys are based here in Vancouver as well), if you have not seen it yet:

Nice going AQVOX for being a shining example of a disgraceful "audiophile" company. And we wonder why audiophiles are often the laughing stock of technology hobbyists? Or why the younger generations with all kinds of other hobbies they may be interested in might not want to take part in "High End" audio?

Here are a couple of stills from the video that epitomizes the kind of product this is:

Notice the fact that this is just a suspiciously "modded" unmanaged 8-port, gigabit, D-Link DGS-108 (currently US$30-40). The fact that they're hiding what they're doing with gluing the screws down and putting goop on components are warning signs even though I'm sure they'll try to justify it with "EMI reduction" or "vibration control" claims. In general I would not recommend audiophiles buy mods unless objectively proven to be beneficial first (including stuff like this).

AQVOX's voodoo Illuminati sticker. Reminiscent of the Peter Belt "foil" stickers (nice article there Greg Weaver - you can still see him doing his "audio analyst" schtick on YouTube) which epitomized the craziness I remember from the turn of the 21st Century, an almost medieval period of audiophilia in my life.

Unbelievable! Don't you love the "Illuminati" stickers?! And some audiophiles deny the encroachment of religious ideology instead of science in products like these? How bizarre. At an MSRP of around €800 (or currently ~US$830 due to still-high USD exchange), this is more than 20x mark-up over the stock D-Link switch! Worth it? Or scam? You tell me.

This is a great example of the kind of companies (yup, let's throw in Synergistic Research here as well) and audiophile beliefs we need to distance ourselves from if we have any hope of sanitizing that term "audiophile". Interestingly, I think AQVOX took down their ethernet switch link already, but this evangelical YouTube review is still up. I guess the voodoo Illuminati iconography "worked" for him - "the difference was clear, reproducible, and rather surprising" he says, sure...

There's also this Positive Feedback article by Dave Clark which leaves me wondering: where do I sign up for the Negative Feedback magazine in which writers honestly are more insightful, evaluate with common sense, and are not afraid to show a little skepticism of voodoo? Look at the Shakti Hallographs in his room pictures BTW, do people still believe in those (another remnant from turn-of-the-21st-Century audiophile insanity)?

Oh yeah, here's one from Martin Colloms of HiFi Critic - come on man, this is ridiculous:

"I would rate the gain in timing, expression, clarity, depth, focus and the more convincing spaciousness, as worth 8-12% for my reference streamer sound."

Rather precise. Why not bump that to a more generous 15% restaurant tip equivalent? Considering the HiFi Critic subscription costs around US$130/year for 4 issues, one would hope the other product reviews are better than this which is no different than the free opinions among the usual websites and YouTube channels! Knowing that Colloms has some technical background, reviews of such poor quality are disappointing to say the least. Is this what it looks like when a guy needs to make a few bucks on writing reviews in the "High End" Audio Industry even without taking in advertising dollars?

[Addendum: As of February 2023, HiFi Critic has ceased publication.]

Anyhow, for us audiophiles, these kinds of dishonorable products sold by painfully obvious snake oil companies are nothing new.

As in the video, I don't necessarily blame audiophiles who buy products like these either because from an audio "culture" perspective, so much of our media and word-of-mouth has allowed the proliferation of these kinds of devices without honest, competent, assessment. When was the last time Stereophile ripped one of these boxes apart and showed hobbyists what the inside looks like as they did in the LTT video? When was the last time the mainstream audiophile press pushed for a blind test, even a simple one like what they did?

Instead, for too long, magazines like Stereophile (with John Atkinson at the helm over the last few decades, and Jim Austin these days) and The Absolute Sound (Robert Harley, Lee Scoggins, et al.) have basically written friendly words for nonsense like this. Most recently, we see the Jason Victor Serinus review of the Nordost QNet switch as another example of this genre of products with no evidence of value. Needless to say, the man liked the US$3200 device - even better with the US$2750 power supply - because apparently Nordost couldn't find an audiophile-sounding wall wart like AQVOX could.

In fact, the ethernet switch was so awesome sounding to Serinus that (emphasis mine):

"The first thing I heard after I installed the QNet was that Gens's [referring to Véronique Gens on Nuits] voice grew in size. Colors were more vivid. As silence filled spaces between notes, the soundstage seemed to expand in all directions. All that from a simple switch?"

Apparently yes, all from that simple (and seemingly fragile) switch. It's all just begging John Atkinson to run measurements on this device or show the putative improvements Serinus said came out from his streamer/DAC. You would think if the difference was so profound, it would not be hard to objectively show a change, right? As usual, nothing was done - not even a teardown. :-|

Thanks "audiophile press" for all the years of "leadership" and providing "education" for consumers.

Like all the nonsense about MQA, this too is a part of the legacy left by the last generation of audiophiles running these publications. IMO, audiophiles these days should try to just ignore or even better yet speak up against these kinds of products and the people who hock them as clearly damaging the reputation for honest hobbyists who want high-fidelity purchased with honest wages.

Looking ahead, what I'm most curious about is, as with the Thomas Paine quote above, will we ever see a time when the audiophile press might find "disposition" or "courage" to address the delusions and errors of a number of players in the Industry? I would not hold my breath with this current generation running the mainstream magazines.

BTW: If you want a high quality awesome router, I can still recommend the Netgear Nighthawk S8000 as a high quality chunk of zinc-alloy ethernet switch. If you can still find one. I bet it sounds just as awesome if not more than the AQVOX and Nordost, plus allow you to tweak things like port priority to favor the precious audio data coming from and to your audio streamer! ;-)

--------------------

Alright guys and gals. Let's drop the blood pressure a bit to end off on something much more down-to-earth and cool.

Recently, Ivan Khlyupin (IVX) at E1DA has been working on updates to his USB "dongle" DACs aiming at improving resolution and he sent me an early production of the latest version, the #9038D6K (based on ESS ES9038Q2M DAC chip) to test out:

Unlike the already available #9038S Gen3 which is a 2.5mm balanced output headphone amp, this identical-looking unit has a single-ended standard 3.5mm headphone jack with very high resolution - close to -120dB THD+N target on a little USB-powered dongle! I've already started listening to it and obviously, I'll have to put this on my measurement rig in the next little while to have a listen/measure/listen myself. ;-)

Also in the same parcel, there's this baby:

That's a prototype E1DA Cosmos Scaler. What this does is provide a low-noise pre-amp front-end to the Cosmos ADC. We can see the dual channel 3.5mm balanced inputs (from DAC to be measured for example) and outputs (to measurement ADC like the Cosmos ADC).

The volume controller is currently pointing to "auto" which will perform an "auto-ranger" function to keep the signal close to a default target of 4.5Vrms by increasing low-level input up to +26dB.

With 200kΩ balanced (100kΩ unbalanced) input impedance, this will deal with one of the criticisms of the Cosmos ADC (variable and lower input impedance). The auto-ranging ability will allow the user to optimize the resolution of the measurement ADC by keeping the measurement level at a higher target (whatever can be done with +26dB amplification). There is also an option to set the output target between 1-10Vrms if default 4.5Vrms is not to one's liking. Ivan has aimed to achieve very fast auto-ranger switching speed as well.

The important factors are of course that the Scaler itself maintains very low noise level and does not introduce any of its own distortions when doing the job which will be at the heart of what I'll be testing out.

Oh yeah, combined with the RIAA EQ ADC firmware as discussed previously, this could also act as a nice +26dB preamp for those doing high quality vinyl rips using the Cosmos ADC.

Off to hack together some 3.5mm-to-XLR balanced cables in preparation for testing... ;-)

Until next time, I hope you're enjoying the music, dear audiophiles!

And to blog reader fgk and those like him in Russia, that there are no new conscription concerns for you - tragic damn waste of lives, opportunities, and resources no-thanks to that Putin regime already.

It's sad that throughout history, the entitled narcissistic passions of a few can create such turmoil for millions, potentially even billions. And this of course is not just an issue with Putin but applies to others across continents and ideologies. Something to think about especially this weekend commemorating Remembrance/Veterans Day in many nations.

Addendum: November 22, 2022

Hey PhilFromTO in the comments. No problem playing Spotify in Windows either to the JRiver 28 WDM driver with the Beelink Mini S:

Playing Spotify music to the JRiver WDM driver into JRiver 28 with HLC VST3 plug-in.

Or we can also use the HLConvolver standalone app to link up audio through VB-Audio Virtual Cable:

No problem with sound quality using the highest Spotify stream which I think is an Ogg Vorbis 320kbps. Maybe one of these days we'll finally get Spotify HiFi... And with any luck, maybe multichannel/Atmos streams.