Using low-power Beelink Mini S as Linux Roon streamer - HQPlayer (PCM 8x, DSD256), and HDMI multichannel. (A superior DIY M-Scaler?)

No worries! Putting a little low-power computer like this on top of the DAC did not affect noise level from the Sabaj DAC RCA/XLR outputs. Don't do this with more powerful computers obviously...

As expressed previously, my philosophy around computer audio is that given the speed of progress in computing technology, it really makes no sense to be building expensive and very fast computers as audio playback/processing devices. For me, if I have a general computer in an audio room, I would want to keep electrical noise low (ie. low power) while also targeting an ideally fanless solution. Inevitably, in less than 10 years, a powerful computer today would be very much obsolete. Monster machines would more than likely end up at the bottom of your closet with little value or interest left even if one wanted to sell. In fact, unless I'm doing a lot of 3D graphics, gaming, or editing videos, IMO progressively smaller, quieter, lower-power (let's aim for 10W or less with excellent performance!), and less expensive machines are where progress is heading for consumer-level general computing.

This is in part why on this blog, I've expressed much more interested in streamer systems like the Raspberry Pi devices or recently silent or very quiet MiniPCs (like the power-limited Beelink Mini S last week). As time goes on, one can judge objectively if technological progress and software updates like new DSP algorithms represent improvements in sound quality. If there are truly gains to be had, the machine can then be repurposed/replaced without feeling that one has wasted hundreds if not thousands of dollars.

At the end of the article last week, I spoke about how I reduced the fan speed and lowered the Mini S power limit to basically "silent" performance. Today, let's consider further what we can do with a machine like that in the audiophile sound room.

First let's quickly recap. What I did was reduced the Beelink Mini S computer's CPU limit to 8.5W with 10W turbo mode (stock CPU TDP is at 15W). This means that when not running at 100%, the machine's total power use would sit between 6-7W idle and 17W max.

With power limiting, the fan speed was dropped down to 30% by 55°C, and raises to 60% by 71°C. By doing this, the machine is basically silent in my quiet sound room (<30dBA SPL in the evenings) and the only noise is when I put my ear against the machine within inches. Ideally, I would still prefer a fanless solution but given the balance of power, energy expenditure, size, and airflow in one's system, a quiet fan is still very reasonable.

Even with the power limit imposed, the speed of the Celeron N9095A is reasonably good:

Notice I highlighted the "Maximum" column in the HWiNFO64 readings. With the fan speed reduced, and ambient room temperature of 21°C, max core and CPU package remains below 70°C after 900 seconds of heavy LinPack processing.

With an average FPU performance more than 23GFlops using LinPack Xtreme, we can put that processing power into audio tasks such as very high quality upsampling as we did using the MeLE Quieter2Q discussed a few months back. However, with newer HQPlayer versions after 4.16.2+ requiring the AVX2 instruction set in Windows, there is no choice but to find a different way to run it on these non-AVX2 Celeron machines... Enter Linux.

Thankfully, Signalyst has continued with keeping the non-Fedora Linux builds free of needing AVX2 which is great news for those of us who are comfortable with installing Ubuntu on our hardware and want to use HQPlayer.

So here's the objective: Let's transform the Beelink Mini S into a low-power, quiet Roon audio streamer based on Ubuntu/Linux, capable of multichannel streaming over HDMI to my AV receiver, and using its processing power for very high quality filtering to my 2-channel USB DAC (Sabaj A20d 2022 currently in the sound room), as well as for DSD transcoding using HQPlayer.

(HQPlayer should be able to handle multichannel upsampling to 192kHz to the HDMI AV receiver as well although I did not try.)

Step-by-step, let's talk about what I did:

1. Go to Ubuntu and download the current LTS build (22.04.1 at this time).

2. I used balenaEtcher to write the Ubuntu ISO to a USB stick:

3. Now we boot into the USB stick on the Beelink Mini S and install the Ubuntu/Linux OS. You can go into the BIOS and tell it to boot from the stick. These days, the Ubuntu install is smart and will allow you to conveniently partition the SSD (this is where the larger 256GB drive comes in useful) so you can run it along side Windows 11.

4. Go into the Beelink BIOS and set the Ubuntu Grub boot partition as higher priority so we can select either Linux/Ubuntu or Windows 11 at computer start-up.

You should have no problems booting into Ubuntu/Linux or Windows 11 at this point.

5. Once you're in Ubuntu desktop, download the Linux build of HQPlayer (current version 4.19) into the machine and install it. The Ubuntu software install app should be able to do this without fuss (right click, "Open with other applications", "Software Install"):

Unfortunately after installing, if you want to uninstall, it's not as easy as clicking on the red trashcan due to various dependencies. For completeness, to uninstall HQPlayer 4 Desktop, do this in a terminal:

sudo dpkg -r hqplayer4desktop 

6. You should be able to run HQPlayer now under Linux on the non-AVX2 Celeron CPU:

As you can see, that's the familiar HQPlayer main window (greyed out) and the "Settings" menu to configure the kind of filtering, dithering, and DSD modulator; yeah, it's complicated and the manual doesn't show in detail what each setting does. I've discussed previously and showed the effects of many of these settings here and here. Be mindful that the program takes time (like maybe 10 seconds) to get itself ready every time you leave the "Settings" menu. This is normal and is seen even on much faster computers. One problem with this delay is that it makes an accurate A/B test difficult due to the limited duration of echoic memory.

As discussed with the MeLE Quieter2Q article, for convenience I like to set the machine to a static IP (eg. 192.168.1.90) which I can easily remember and point to whenever necessary.

In order to link HQPlayer with Roon, turn on the network control icon:

7. Now point Roon at the HQPlayer machine (ie. my static IP):

And you should see HQPlayer among your audio output devices:

8. While I'll be using HQPlayer as my high-quality 2-channel upsampled stereo device to the Sabaj DAC, let's turn our attention to the multichannel HDMI Roon output to my Yamaha AV receiver:

A. Install Linux Roon Bridge which is also needed if you just want to stream to USB without HQPlayer. Follow Roon's instructions here. I just did the "Easy Installer" route which worked without any problems. I appreciate Roon's detailed, easy to follow, documentation. [Note, there could be permission errors with Curl - make sure to install it in the command line with "sudo apt-get install curl" rather than through Snap Store in Ubuntu.]

B. Set the Ubuntu output to the multichannel layout in the "Sound" settings menu. For me, I went with 7.1:

C. Within Roon, you should be able to see the HDMI outputs. Looking from the back of the Beelink machine, the rightmost HDMI port is what I'm using, this is mapped to "HDA Intel PCH HDMI 0" which I can rename as "Beelink-HDMI 7.1":

In the "Device setup" for "Beelink-HDMI 7.1", we can adjust some of the details for the multichannel output including the channel layout (which correlates with the setting I have in Ubuntu) and you can send stereo/mono as 7.1. I noticed in Windows, I had to turn this to "Yes" otherwise sound would not play. Linux does a better job and this could be set to default "No" which then allows the receiver to process 2.0 input as it sees fit. For example, most modern AV receivers can take the 2.0 and remix it through Dolby Surround Upmixer (DSU). Doing this fills out the sound by steering some content to the rears, surround, center and Atmos height speakers which I often enjoy. Sometimes I prefer DTS Neural:X upmixing, but it tends to be a little too aggressive with surround and height channels than DSU.

With that, you should be all set with the HDMI output!

9. Consider Linux power options and turning on auto login. I've noticed that when Ubuntu puts the machine to sleep and then wakes up (eg. with movement of the trackpad or key stroke), it does a really good job of picking up the connected HDMI AV receiver and handshaking with the USB DAC. This is great because I like saving power so it means I don't need the machine turned on 24/7.

A few things to adjust to make this work smoother:

Make sure to turn screen blanking off to prevent the HDMI output from shutting down unless really going to sleep. And the "Automatic Suspend" function which puts the machine to sleep I've set to 120 minutes without activity. The computer will turn back on when I touch the keyboard which is equivalent to a remote power button. [If you need longer than 2 hours before machine suspends, follow this discussion and use dconf-editor to change the value with your arbitrary amount.]

Also, for convenience after a power reset, so you don't have to log back in, you can turn on "Automatic Login":

And so long as you're not too concerned about security of the machine (even if hacked through my LAN firewall, there's no sensitive material nor much on this computer a hacker can likely do other than screw up the audio installation!), turn off the "Lock Screen":

10. Finally, have HQPlayer 4 Desktop start automatically when booting up. This is again for convenience since I'm using this box basically as an audio streaming appliance. If there's a power outage, the machine will restart into a state where it's ready to stream:

Go into "Startup Applications Preferences" and add HQPlayer with the command "/usr/bin/hqplayer4desktop" which is where the software is installed by default.

After these steps, you should be well on your way to playing around with HQPlayer's myriad settings as well as HDMI multichannel streaming through Roon. It should all work quite smoothly "headless" (ie. you should not need to turn on the TV/monitor, just use Roon on a phone or tablet to control). The only caveat is to make sure the output devices like the HDMI receiver and DAC are turned on before you press play in case it throws up an error window about device not present; if this happens you'll need to turn on the screen for a look at what happened.

First, here's a look at multichannel streaming to my AV receiver:

Notice that I'm running a 2-channel 66k taps convolution FIR filter for some correction to the front speakers. At some point, I'll run a proper measurement for full multichannel room correction.

Notice how little stress is being placed on the Celeron when streaming 5.1 multichannel from Roon:

Barely 5% CPU utilization across the cores with a 5.1 32-bit/48kHz stream. Even if we stream 7.1 32/192, we would not be saturating the power-limited Beelink Mini S.

However, we do need more CPU power with fancy filtering using HQPlayer...

Some suggested HQPlayer settings using the quiet/power-limited Beelink Mini S

Okay so on the HQPlayer side, what settings should one choose knowing the processing limits of the Mini S' Celeron N5095A CPU? Let me show you my two "Audiophile" settings that sound great on the system here:

As you can see, for PCM upsampling, I like using "sinc-L" combined with "LNS15" as the dithering algorithm. This combination when upsampling 44.1/48kHz material 8x to 352.8/384kHz uses an extreme "megatap" (1 million taps, because we can :-) filter length that results in an essentially ideal "brickwall filter" with optimal frequency response and transient speed in the time domain.

On the SDM (DSD) side, I'll use a combination of the "poly-sinc-long-ip-2s" oversampling setting which is an intermediate phase option with low pre-ringing and less phase shifting than pure minimum phase, combined with a very good 7th-order noise shaped modulator in the form of "ASDM7" (which I had tested before), converting the PCM to DSD256 (44.1kHz x 256, the Sabaj A20d 2022 with its XMOS processor is unable to handle 48kHz x 256 DSD currently). The little Celeron CPU is capable of these settings without issue as I'll show in a little bit.

Beyond the 1x sample rate, "Filter Nx" and "Oversampling Nx" are IMO not important, just throw some lower-performance settings in there to save CPU cycles and it'll be great. The reason is that when we're dealing with 88.2/96+kHz material, Nyquist frequency is already up at 44.1/48kHz so whatever filtering is being done up there will obviously be well beyond human hearing. Anything decent will do ("poly-sinc-short-lp(-2s)" is excellent).

Finally, notice I have the "Vol Max" set at -1dB for a bit of headroom in case of intersample overloading. Since I almost always listen with volume normalization applied, this is just an extra level of assurance.

How well do these HQPlayer 4 "Audiophile" settings work? Let's have a look at the "Digital Filter Composite" (extension of the Reis Test) from the Sabaj A20d 2022 DAC XLR output:

--- "Ideal Megatap" setting (PCM 8x 352.8/384kHz, sinc-L + LNS15) ---

XLR output and noise levels I get with the Mini S sitting on top of the Sabaj DAC.

Beautifully clean filtering with obvious "brick wall" right at 22.05kHz (44.1kHz sample rate). No intersample overload. Very low harmonic and intermodulation distortions. Very low noise level. We can zoom in a bit into the 10kHz to 50kHz portion to have a better appreciation of the steepness of the filter:

Beautiful; way beyond the resolution of any human ear and as good as the Chord M-Scaler, without the excesses of PGGB. ;-)

For completeness, here's a look at the 16/44.1 impulse response with this setting:

No surprise. We're looking at a very long, linear phase impulse response to achieve the steep filtering and strong stopband rejection.

--- "Audiophile Choice" DSD setting (DSD256, poly-sinc-long-ip-2s, ASDM7) ---

Now let's use the DSD option that also performs beautifully. But we'll throw in an intermediate phase setting aimed at reducing pre-ringing that can happen when reproducing modern content with peak-limited, even clipped waveforms.

DSD256 is a natural target sample rate as it pushes noise well above audible. Here's the wideband DFC from 15Hz to 192kHz using the "Audiophile Choice" DSD256 parameters:

Again, that looks great. While there is increased ultrasonic noise due to the 7th-order DSD noise shaping modulator, the rise occurs above 100kHz. If we zoom into the 10kHz-50kHz portion to have a better look:

We can see the shape of the roll-off steepening after 20kHz characteristic of the "poly-sinc-long" filter, well attenuated by 22.05kHz. Since I chose the "-ip-2s" variant, here's the intermediate phase impulse response:

By allowing only a tiny amount of pre-ringing when we run into poorly low-passed content, we also reduce the phase shift compared to a minimum phase filter. We can measure this difference by looking at the phase characteristic of a frequency sweep comparing linear, intermediate, and minimum phase "poly-sinc-long" settings:

Intermediate phase setting staying +/-30° up to 18kHz, and at most -90° by 20kHz; very benign.

With the highest quality "audiophile" recordings, I'll happily use the "Ideal Megatap" PCM 8x setting. For rock / pop, especially "Loudness War" stuff over the last few decades, I'll happily play those with the "Audiophile Choice" DSD256 setting to reduce pre-ringing in the conversion. While intellectually I guess I can make an argument for this differentiation, these are thoughts of an obsessive audiophile fooling around with minutiae; the sonic differences IMO is subtle at best. ;-)

Suppose we dispense with technical "high-fidelity" accuracy and let's consider a very nice sounding combination of settings that could subjectively "open up" the sound stage, and can even excite the sense of "presence" during playback.

I reported years ago when I first explored PCM-to-DSD conversion with JRiver that with A/B listening, DSD can sound different than PCM. My suspicion is that many listeners enjoy SACD because of the fact that it uses DSD64 rather than the higher DSD128/256+ rates and that the higher we go with DSD samplerates, the more it just sounds like clean PCM.

DSD64 introduces noise into the system which in certain instances I believe imparts a subjectively pleasant effect that to my ears sounds like a subtle expansion of the soundstage. This is perhaps similar to vacuum tube amps or LP playback where it is because of the imperfections, some listeners will report more subjective enjoyment, rather than suggesting that technologies such as these represent higher fidelity or higher resolution.

These days, as we achieve objectively amazing levels of fidelity, it's this "euphonic" effect that IMO has become more interesting. As an exercise in self-awareness, it would be great if we audiophiles can apply objective analysis to identify for ourselves which types of distortions we like in the playback!

--- "Euphonophile Choice" DSD setting (DSD64, poly-sinc-xtr-short-mp-2s, ASDM5) ---

Here's one that I've been playing around with recently and sounds really good to me:

This is how the DFC looks using these filter and modulator options converting 16/44.1 to DSD64:

Within the audioband up to 20kHz, the performance remains excellent. Obviously we're also seeing the ultrasonic effect of DSD64 noise shaping rising above 22kHz. Here's a look at the impulse response for such a setting:

Captured at 384kHz; bandwidth to 192kHz.

Notice the use of minimum phase which some listeners subjectively describe as pleasant (although previous blind testing doesn't seem to show strong preferences), and the low-level high frequency noise can be seen in the tracing. In fact, we can see the DSD modulator noise in a 0dBFS 1kHz sine wave:

"Imperfect" but "analogue" looking. ;-) That is the promise of the SACD sound, right!? And doesn't that also look much better than stair-stepped "digital" PCM NOS waveforms?!

HQPlayer CPU Demands on Celeron N5095A

With the low-power Beelink Mini S, obviously we can't expect to play with all the DSD options HQPlayer provides. In fact some of the settings would bring mighty Intel i9 and AMD Ryzen 9 CPUs to their knees! DSD512 oversampling would also benefit from GPU processing as previously tested (CUDA offload to my nVidia GTX 1080).

With the settings above, let's see how much of the Celeron N5095 is being utilized when playing music real-time - here's the CPU performance monitor when playing a minute of music using the settings:

Playing various albums through the filters, correlating with CPU load - Swing Out Singles, Elvis OST, and Lover. Great sound quality with excellent soundstage and precise transient attacks across the board. Maybe more "euphonic" with the DSD64 playback.

As you can see, "megatap" PCM 8x upsampling is no big deal, requiring <10% CPU load to upsample to 352.8kHz. Yeah, the computer would have no problem upsampling this to 16x 705.6/768kHz and the Sabaj DAC will handle that as well, but there's simply no need. When it comes to DSD, more CPU cycles are needed for DSD256 compared to DSD64 processing. So long as the CPU threads don't hit 100%, there's no deterioration in the sound quality. With this Celeron processor, most of the 2-stage "-2s" oversampling settings work well and one can experiment with other modulator options as well (be careful with those "EC" SDM variants, I personally don't hear/measure much if any difference using them).

Again, don't worry about the upsampling setting for "Filtering Nx" or "Oversampling Nx" when sending high-res sampling rates >44.1/48kHz. Just use a setting that doesn't put too much strain on the CPU like this:

As you can see, when sent over 192kHz material (demo tracks from Barry Diament's Soundkeeper Recordings), HQPlayer will use the less demanding "poly-sinc-short" setting when upsampling to DSD256 and this takes slightly less processing than the higher quality "poly-sinc-long" when converting 1X content with "Audiophile Choice" DSD256.

Summary:

Beelink Mini S - streamer, processor, playing DSD64 "Euphonophile Choice" setting to Sabaj A20d 2022 DAC.

There you go! An inexpensive Beelink Mini S computer transformed into a low power (less than 17W peak), very acoustically and electronically quiet machine that can handle Roon multichannel 7.1 streams through HDMI as well as high quality filtering through HQPlayer to a good USB 2-channel DAC in Linux/Ubuntu OS.

HQPlayer Desktop costs around US$215, plus US$180 for the Beelink Mini S machine, so in total for around US$400 plus a bit of DIY time & know-how, we can achieve "megatap" PCM upsampling easily equivalent (and actually superior) to the Chord Hugo M-Scaler which they're asking US$5000+ for. Furthermore, HQPlayer is more versatile with many filtering algorithm options and the ability to convert PCM to DSD adds another dimension of listening/experimentation.

Again, let's not forget the ability to stream multichannel content. This is what "value" looks like! Yeah, I know, Roon and the music server computer also cost money. Plus you need at least a good amp and some great speakers to gain from all this processing. Nobody said audiophilia is totally without some base cost even if we keep the cost of parts like the computer down!

To end, I want to address some of the reviews I've seen about the M-Scaler like this one (from August 2019 - I hate it when articles don't include a date), or this obviously uncritical multi-part piece in praise of the M-Scaler from 2021.

In that 2019 article, the author claims (emphasis mine): 

"I tried the SMSL M300 with and without the Hugo M Scaler and the difference was huge, the M Scaler improved the sound in quite an incredible way. The level of improvement that would make you buy one within a week of testing it exactly the same as me (sic)."

Seriously guys, talk about unbelievable superlatives spoken in ways that suggest the reviewer really has no idea what this thing technically does! If one's DAC had very poor filtering or no filtering at all (ie. NOS), then sure, adding something like the M-Scaler could be significant. Otherwise, permit me to be blunt and say that with a competent modern DAC even like the slightly older SMSL M300, the difference cannot somehow be "huge" or "incredible". Even more so since he's connecting the M-Scaler to the SMSL DAC through S/PDIF (see the video) and therefore basically just upsampling to 192kHz through that inherently more jittery interface than the USB port! If an audiophile seriously wants to hear "huge" improvements, make sure to try some DSP room correction.

I see Amir of ASR has a YouTube review of the Chord M-Scaler (and the write-up). Very interesting objective findings suggesting that the supposedly "incredible" change in sound could just be due to higher noise and jitter created by the M-Scaler?!** Anyhow, all I can say is that the HQPlayer with the "Ideal Megatap" PCM setting shown above performs better than what the Hugo M-Scaler is doing in that video/review. Of course realize that when we're just upsampling, there is no improvement in resolution of the DAC, all it's doing is ideal "brickwall" low-pass filtering. Also, it looks like the M-Scaler has poorer stopband attenuation, or maybe suffers from overloading with strong signals. Yeesh, given the typical Chord talking points about "big number" taps and bits of resolution, talk about a disappointing finding!

As for HQPlayer, my comments/suggestions from a few months back remain the same. As you can see, the Celeron without AVX2 works just fine (especially with PCM, and I'm glad the Linux version is still compatible). I still wish the selection of filtering algorithms can be simplified. For "core" settings like "poly-sinc", it would be nice to be able to fine tune the steepness with custom number of taps, cutoff frequency, and phase settings.

I actually find the vast majority of settings in HQPlayer "too accurate" (and basically sounding the same, subjective claims of hearing massive differences between them I suspect is placebo) especially when I specifically want to aim to try intentional suboptimal "euphonophile" experimentation. Just as it would be silly to claim that the Chord Hugo M-Scaler made an "incredible" difference, it would be unreasonable to claim that HQPlayer "poly-sinc-gauss-long" sounds significantly different from "poly-sinc-ext"; the differences are demonstrably minor and mainly ultrasonic in my testing!

Instead of yet more "accurate" settings, I think it would be more fun to implement something like the PonoPlayer's "Ayre Listen" weak filtering with DSD64 oversampling using the 5th-order modulator. As far as I can tell, currently this is not possible in HQPlayer because all those options are too darn "good".

Anyhow, have fun folks! Lots of geeky opportunities these days in the world of computer audio to do interesting things with the signal even if objectively the differences are not major (whether one actually hears much of this is of course another discussion).

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

To end, I think as audiophiles it's fun to find ways of optimizing devices to propel fidelity forward instead of spending money on things without clear knowledge or any evidence of actual benefit.

Given how good playback fidelity is these days with much of the modern reputable gear, I think it's becoming more important as consumers to ask (if not demand) that device manufacturers show us just how much better their new machines are objectively. This is important if the cost appears exorbitant; in this way, I guess my core philosophy is the polar opposite of this guy who seems to think it's a good idea to buy first, talk about "feelings" and "impressions", but shares little actual information or technical details about the products.

I must say that even though he doesn't really have a "lab" (by definition one performs science in a laboratory, this is not science), the room doesn't appear spacious enough for the kind of gear he's using, and I'm sure the Boulder amps must sound great, I am mesmerized by his sincerity (including this recent transparent discussion of a flip-flop), and the nature of the character as he bloviates about the benefits of his level of "experience" (check out 7:30). Other than the Boulders, if you're interested in big, black, impressive looking, powerful monoblock amps, I would also suggest auditioning the German-engineered mbl's (which would also go well with his Stromtank - see video).

BTW, I noticed that the Boulder 3050 uses a variable bias circuit that adjusts power level to keep current utilization low and maintain Class A when encountering stronger transients. I'd be curious how it handles the first few cycles of something like the TIM test signal and while at it, the device's damping factor / output impedance across the audible frequencies would be interesting to see. The company has a technical brochure on the 3050 amplifier for those curious about some of what it does.

Hope you're all enjoying your audiophile journeys. And of course the plethora of music along the way!

** Perhaps not unfair to give Rob Watts the benefit of the doubt. Maybe he knows that the Hugo M-Scaler adds noise and jitter, and this was intentional!? Just like the "Euphonophile Choice" DSD64 setting is meant to add noise by design. Maybe he knows that we know that some audiophiles prefer those "euphonic" imperfections but can't say it openly based on the Chord reputation of technical perfection. ;-)

Assuming it's even audible in a level-matched listening test, I suspect the truth is that without those anomalies, the 1M-tap M-Scaler would sound no different than the already-ultra-steep filtering of even something like the old Chord Mojo 1 with its "mere" 26k-taps filter.

Addendum (October 23, 2022):

In terms of "headless" control of HQPlayer, note that we can change the filter settings using apps like HQPDcontrol v4 for Android and iOS. Just connect the app on your phone/tablet with the IP of the Beelink computer (remember, static 192.168.1.90 for me above), and you can switch between your PCM/DSD settings, and fine-tune whatever parameters you decide when the music is paused.

Works well. Nice and convenient.