Building a DIY Music Server

[Energy]

@StreamFidelity

 

Yes, you can. I recommend using a GX16-2 chassis mount male connector attached to the back of the case nearby the EPS connections with four 16AWG wires running from the GX16-2 to the 4-Pin EPS. Two wires for positive and two wires for the negative. Strip the wires with a wire stripper and crimp it to a gold plated female dupont 2.54mm connector that is made from copper (most are brass).

 

The crimp tool, dupont connector, and EPS housing (8-pin) in black or white:

 

 

The more wires used lowers the impedance thus decreases the chance of voltage sag which ultimately improves voltage tolerance.
 

What I do is crimp the wire to the dupont connector, add some flux, and quickly solder the two together to create a better conducting joint. The teflon insulation should still be grappled by the bottom part of the Dupont connector with a 1/32 heatsink going over it to help add flexibility so when the wire bends after it is inserted into the plastic EPS shell, it can move freely without ripping into the wire’s strands or it’s insulation. For those wanting to add JSSG360 shielding, I recommend putting the 16AWG wire through a small cotton tube or paracord type I before adding the shield as this adds physical distance which helps reduce cable capacitance. Both material help with the accumulation of tribolectric charge (static charge) although cotton would be more optimal than the nylon found in paracord since cotton is considered neutral while nylon by nature has a negative charge.

 

My question for @Nenon is whether or not it is possible to use two different 12V power supplies of 10A if there are two 8-Pin connectors used for EPS or will the combination cause a short?

[Energy]

On 2/20/2020 at 10:48 AM, auricgoldfinger said:

I've been trying to determine the quality of the clock offered by Fidelity Audio, and it has been a struggle.  Apparently, I am the first person to have asked about the phase noise characteristics of their clock.  It seems they may devote more effort to their voltage regulator than Pink Faun. 

 

This 40hz data in the attached graph is from the supplier of the range of oscillators they use.  Any thoughts?

 

phfobkoogonohchd.bmp 271.54 kB · 19 downloads

 

 

 

I've come across them before but the fact that they create clock boards but do not have phase noise plot graphs shown is otherworldly hence I never messaged them.

 

At 100Hz it looks around -108dB. Even a Crystek CCHD-575 at 100Hz can do -121dB or even -130dB depending on the frequency.

Powering them with low noise LT3045's with additional regulation (more added in series) will help PSRR which in turn should lower phase noise further. There are clocks out there that are even better, like Pink Faun Ultra OCXO which can do -145dB at 100Hz, but you gotta pay for it. These clocks however do not perform at the level of masterclocks like MUTEC REF 10, SOtM sCLK-OCX10, or Cybershaft Ultimate. These masterclocks although do not come in the frequencies required for direct connect, have 100Hz phase at around ~147.5dB but it's their 1Hz phase noise that matters. For the Pink Faun at 1Hz it's -100dB whereas the masterclock's (like the one listed above) are around -115dB. Companies like Cybershaft and MUTEC have even created custom models that cost a whole lot more for those who want better performing SC cut crystals (oscillators). For example, Cybershaft has their OP21 model that can do -121dB at 1Hz (hence the name) and MUTEC has responded in the same fashion by releasing their SE-120 model.

 

My guess on why 1Hz is so important is that improving it improves the total jitter of the system by a good margin. There are some audiophiles (eg. Cybershaft OP20 vs OP17 comparison) who can hear a difference in a few -dB's at 1Hz. My assumption is other than the total jitter improvement, 1Hz is fairly close to the 0Hz DC that is provided by the power supply which in turn affects it a lot more than other frequencies. But who really knows 😵

 

Here are some examples that can affect phase noise:

1. IEC with EMI/RFI Filter

2. Encapsulated Step Down Transformer (perhaps even balanced wiring) to limit magnetic radiation

3. Low ESR capacitors for quicker high pulse currents to the oscillator crystal and noise filtering

4. Oven to heat the crystal so it's frequencies are more stable

5. Galvanic Isolation between outputs in order to not influence the crystal

6. Proper use of film capacitors and star grounding to shunt noise and leakage currents from influencing crystal

7. Decoupling of the chassis with something like IsoAcoustics Gaia III to decouple it from the earths natural movements/shifts of it's tectonic plates

8. Shielding or using EMI/RFI Absorbing sheets to prevent high frequency EMI/RFI from affecting the crystal

 

Both (7) shielding and (8) decoupling might also help with the natural 7.83 Hz frequency the earth makes. It's a lot $&#%.

 

So if you are looking for a spot on replacement, you can just use Crystek CCHD 575's in the right frequency. If you want a little more, then I'd use a a small linear power supply connected to a LDOVR LT3045 to power the oscillator. However at higher levels, you'll want to read into the phase noise plot of each oscillator that you are thinking of buying.

 

For this reason is why I went with a SOtM sCLK-EX clock module and use a MUTEC REF 10 connected to it for even lower phase noise. This is thanks to the masterclock's low phase noise frequency now being used as a reference for the sCLK-EX. In turn it produces clock in whatever frequency you desire but with lower phase noise than if it did not have a reference connected to it. It is pricey but comes with four outputs. Each output costs $50 to program. The benefit of this method is for those who require more outputs and want to use a simple solution. It's also great for those who have already planned to get a masterclock in the future. I have one in the music server and another in the tX-USBultra which connects to all the clocks that are in the endpoint. I plan on getting another one and putting it into the Holo Audio May so it can replace the main clock (Crystek CCHD-957) and those used for multiple FPGA's. This in turn creates the lowest overall phase noise from when the music leaves the audio system until it's arrival at the DAC before conversion to analog. It in turn helps to create produce the most analog sound since it ultimately trying to remove the digital problem (jitter). The only sacrifice with this method is a lot of "spaghetti" clock cables dangling that can affect phase noise if it isn't short enough or using a quality cable. 

 

All in all, when it comes to clocks you must know that there are some clock locations that just simply cannot be improved. For example, some high quality DAC 's like Denafrips Terminator or Holo Audio May do not come with masterclock inputs. They rely on Crystek CCHD-957 that is used in their product as a reference for the PLL to create other frequencies out of. So even if they are using PLL like with other clock synthesizers that accept masterclock frequencies as a reference... the only way for it to accept a 10MHz signal from the masterclock directly is if the PLL chip within the DAC is reprogrammed to accept 10MHz. I don't think this works like wide voltage tolerance NUC's that accept a certain range of voltages and adjusts accordingly but correctly me if I'm wrong.

 

On 2/20/2020 at 3:17 PM, dminches said:

I would need a different motherboard to use PCI storage since my Supermicro mobo only has 2 PCI lanes and they are being used by network and USB cards.

 

Is there an ATX motherboard which uses ECC ram?  I would like to continue to use the Apacer RAM, if possible.

 

 

Intel server motherboards with Xeon work with ECC ram. 

AMD motherboards do too and does not require a server motherboard. They work for consumer motherboards and CPU's (eg. AMD Ryzen 2 and Ryzen 3). The main benefit of the server is that there is more slots to put ram which increases the bandwidth. This in some cases like the Taiko SGM Extreme improves sound quality. Having multiple CPU's however requires RDIMM's. On each RDIMM memory module there is a separate IC used for buffering (processing) to organize the workload given by multiple CPU's. With a single CPU only UDIMM's are needed as they are unbuffered (hence the "u") so whatever tasks comes comes and gets sent to DRAM. The use of these registered (hence the "r") buffer chips lighten the workload of the CPU memory controller so they work less hard and might improve the latency and ultimately the sound quality. One drawback of using multiple sticks of ram for a setup like this is that the more sticks used, the more voltage tolerance problems occur. This occurrence is from sticks that have memory module IC's that come from various batches used to make the final RAM stick. Thus if one really wants to increase bandwidth, then it's recommended to use Industrial RAM which solves this problem as they come from the same batch and are created with closer tolerances. Furthermore it is known that ECC sounds better as it corrects on the soft/hard bits that computers tend to make. Some people on the forum has found that to be the case including myself. So between Intel and AMD, I am more lenient of the later. However I believe romaz or austinpop found that their AMD machine sounded more digital than their Intel one unless they used a JCAT Net Card. So do experiment and let us know your findings.

 

Lastly, for those who are using HDPLEX 200W + HDPLEX 400W/800W HiFi DC-ATX or 800W. I think the idea is ridonculous. You are essentially using a 3mV (3,000µV) device and adding it with a 10mV (10,000µV) device to have a max output noise of 13mV (13,000µV). The HDPLEX 400W ATX Linear Power Supply may be a failure in it's independent output but it's ATX modular Output is a heck of a lot better:

 

Just saying but 3mV beats 13mV. And look at that DC current headroom!

So instead of upgrading to the 800W version of the DC-ATX converter for nearly $200 more just to gain a few decimals of an improvement in voltage tolerance, the money can probably be better invested towards the HDPLEX 400W ATX Linear Power Supply, or if you know circuitry, use the money to build a linear power supply out of the Belleson SPHP for <5µV on it's output instead of 3,000µV.

[Energy]

To clarify, the end goal has always been sound quality, which is subjective but at least that'll start us with some commonality. 

 

Everything requires proper implementation. Even with proper implementation sometimes certain components can sound poorly when used in a system that has bad synergy. Which is why this hobby is a complete loop and in order to understand the system better, more specifically YOUR system, you have to change out things one by one while making sure everything is properly burned in so there is no placebo going on. Which is why I envy your persistence for this hobby and enjoy your updates.

 

Speaking quickly on the RAM. I found that frequency does not play a big of a role than say industrial quality or ECC specified RAM, hence why the 3200MHz consumer sticks I was using too did not sound as good. This also includes the CL timings. The Industrial/Wide Temperature ones have shown good results with those who have tried them whereas ECC is still only a recent matter. It does correct the soft/hard bit errors but you just might not hear it. Or you might hear something different but not necessarily better in this case. It requires either a very sensitive setup, the right headphones, amplifier, DAC, or the right amount of sleep. After all, our eyes and ears work better on different days. To say so conclusively as to what a wise choice would be is that no one should build a system around ECC "enabled" RAM, that I'll agree with you 100%. But the difference is there. Either from ECC or just the quality of the module themselves. You yourself noticed this when they were used without ECC enabled in another system. I'm not fond of taking guesses but my best one would be that ECC/Industrial produced RAM come in even smaller batches than industrial alone ones so perhaps their tolerances are even closer.

For those looking to upgrade their RAM, Industrial ones SHOULD give you an improvement (due to closer voltage tolerances between IC's), but don't fixate too heavily on ECC RAM. In this hobby of wishful thinking even I must admit what I heard in improvements could be from wishful thinking and buyers justification mindset, all forms of placebo. Regardless of which, even in that category, at the end it's all about self satisfaction may it a fairy tale. After all, we're all trying to wake up like The Matrix here and see what the truth really is, hence all the experimentation. On the subject of RAM, currently I have purchased an aluminum heatsink for the four sticks of Apacer I have. I am going to see if cooling them helps in any way. If not, maybe the metal shielding will do something in rejecting EMI/RFI so they can perform better, it's a valid thought right?

 

What you said on power supplies is true. It's not all about output noise but a collection of things that goes on the topic of implementation yet again. Things such as voltage tolerances, current headroom, output noise, transient response, discrete vs non-discrete, super-capacitors vs without, amount of regulation, Low Pass LC Filters (Pi, Legendry, Chebyshev Pi, Butterworth) vs basic filtering, balanced step down transformers, ferrite/chokes (wrapped around wires), boutique parts (rhodium IEC, GX16 connectors, audiophile wires and capacitors, EMI/RFI absorbing sheets,), proper cooling, and chassis decoupling all play vital roles in how well a power supply performs and which would be more optimal for what circuit. I only list the 10 + 3 = 13 as a way to show that we are back paddling. Audiophiles who are known to use inefficient linear power supplies that are in the single digit uV and now paying little attention to noise level and utilizing higher efficiency but noisy SMPS converters. It's so opposite. There are better options. In general higher current, lower noise, and voltage tolerance are the three main staple. The boutique part is subjective (although I am guilty on leaning here) and the methodology of how the power supply is built affects each line-load differently, not just because their impedance or current draw differs but their preference for power can variate. Like how the input stage of some amplifiers having high impedance, the output impedance of power supplies being low might also not always be a good thing. A transient response that is delayed might be better than a faster one for a particular load. Maybe the load can't appreciate high pulse currents. The list goes on. It all comes back to experimentation but more importantly, know what you are experimenting. For example, Farad uses a Pi filter to create a cutoff frequency to not allow high frequencies (eg. cutoff after 0Hz to not allow noise to pass other than the DC alone) to go through and into the voltage regulator. Because they incorporate this, I have found it to make more of an improvement on digital gear as those gears can appreciate that improved shunting of high frequency noise. Just my 2 cents. Know the inner workings of things so you can experiment in all directions. Don't try the same place twice unless it is somewhat different or if you want to confirm one final time before departing with a set conclusion.

 

I was unaware that it was easier to make a low phase noise OCXO at 10MHz than it was to make it in 25MHz, thank you. I still believe the phase noise matters but like you say, it only gives us a small window into what it's really capable of (good and bad). Regardless of which clock, the layout still stands that it should have the shortest path possible to where it is being connected to. This can be a tough challenge especially like your photo when finding a properly location to replace the chipset oscillator. I for one am at the other end of you. I've tried a bunch of Crystek and custom measured NDK's with even lower noise before listening to @elan120 and switching over to the sCLK-EX. It was yet still better than everything I had tried and even more so with a masterclock despite a longer run. I hope to meet you half way one day by replacing one or two clocks with Pink Faun Ultra OCXO in the most vital area of the system just to see if there is any difference between the two to justify the purchase of such an expensive OCXO. But it needs to be done to really not look back. I think you feel the same in regards to the SOtM clocks. And don't be shy of your belief. It's been said by Cybershaft and MUTEC that a clock without vibration isolation can increase by a couple +dB. Hence why I prefer having the MUTEC REF 10's OCXO separate from the rest of the system, not mounted on shaky mounting brackets, and on isolation feet's like the IsoAcoustics Gaia III. Also you never know how much light movement the earth makes as it shifts it's tectonic plates and crates a light whiplash on the audio gear or how it's natural 7.83 hertz frequency can affect the SC cut crystal. 

 

Either way, I firmly believe what you're doing by starting with the best quality OCXO and working up from there is a smart move. I on the other hand populated my whole system with sCLK-EX. I may get the musical benefits now but hearing a clock swap is much more difficult although I will try. It's too late to throw back in old oscillators. You starting from stock and then jumping end game would surely be a dramatic change enough to figure out what location has the biggest perks. I await for your experiments.

 

Lastly, I have some carbon pipes in many different sizes. If you want to do something with them just let me know the size and I'll cut it down for you and mail it to you. It could be put to good use. Maybe around a wire for shielding purposes and dampening as you say (true). I also have a roll of  tinned copper braid a 1000M roll of mylar aluminum foil in 17mm (I think..). 

 

 

  

 

[Energy]

1 hour ago, TheAttorney said:

A great post that answers so many questions, but I can still think of a few more on RAM 🙂:

 

1. Is 2 better than 1? E.g. is 2 x 4GB better than 1 x 8GB?

(Some say 2x is better because of load balancing. OTOH it may result in slightly greater power consumption and noise. And greater cost). 

 

2. Is it a bad idea to mix 2 slightly different memory specs?

In particular, a 2400 Apacer and a 2666 Apacer in adjacent slots.

 

3. How big a SQ step is the 2666 above a 2400 Apacer, compared to a 2400 Apacer above a typical non-industrial RAM?

And same question for ECC above non-ECC?

I'm guessing there are diminishing returns here.

 

Thank you

 

Starting off are general answers, but the answers themselves are open to variable change depending on your system. 

1. In the multiple test configurations that I ran, 2 sticks sounded better than one. But only when the device was given adequate power. That means that on a music server that is given enough current, the dual channel bandwidth (or multiple sticks) and latency of data queue’s improved the sound quality despite requiring another 1.2V memory module that added additional power consumption. The closer the PSU was to it’s max current limitation, the more narrow these benefits became. For example on the endpoint I once used a 3A linear power supply that was cutting it close and when the second RAM was added, it actually became harmful to the sound. It wasn’t until I had switched over to a 5A did the improvement of the second stick become apparent. YMMV.

 

2. If they are of similar build then running both Apacer model’s at the same frequency shouldn’t be too far off, given that their PCB layout (traces) and IC’s are similar. However if the comparison was with different brands or models then given the two reasons above it would be a bad idea to mix them. Even Apacer sticks that run the same frequency but designed with different initial stated frequencies can come from different batches that don’t work as similarly. Given that, it’s probably a better idea to get only industrial with the same frequencies to be more certain that they perform more similarly. But right now harmful it is with the current configuration you have really depends.

 

3. In very rough numbers, I found the 2666 to be maybe 2% better than the 2400 (both industrial), and ECC was only maybe 1-2% at most. Both were at least 3% better than non-industrial. The changes in these specifications was more significant for the endpoint than it was for the music server. With the endpoint, since I use an AMD based NUC/Mini PC called the ASROCK iBOX-V1000 that supported ECC out of the box, I started off with low CL14 HyperX consumer RAM before directly moving over to ECC/Industrial. Where I got maybe 7% total improvement on the server, the endpoint on the other hand felt closer to 10%. My guess was that it is in the middle Of the audio chain and responsible for handing off the music file hence played a slightly more important role whether that operation went smoothy or not. An operation where RAM plays a fairly important role since the unit is made up of very few things (motherboard, CPU, and memory). Not to mention the AudioLinux operating system was off-loaded to the RAM via RAMBOOT so that could be a major contributor on why RAM made more of a difference here.

 

I’m sure Nenon will chime in on this as well with his thoughts.

[Energy]

2 hours ago, RickyV said:

Have you compared the NUC7I7DN.. and the iBox v1000 and which did you prefer? 

 

Yes I have. I haven't shared the news with anyone else other than with friends but it does sound better to me. I've got some of my audiophile friends to switch over to it. Some from NUC7I7DNHE's while other's from UltraRendu. To start off, aside from the sound quality, it has some good hardware. It's CPU and GPU (it's an APU thus combined) is a little faster than the Intel while it it's memory supports ECC and has a higher limit (3200MHz vs 2400MHz). It comes with a display port in case you ever want to use it as a normal computer and wish to pair it with higher refresh rate monitors. For the same price as the Intel offering you get better performance which when running it as an endpoint, the performance affects the sound quite a lot.

 

The setup I have is Music Server > EtherRegen > Endpoint

 

The endpoint runs AudioLinux via ramboot. The memory modules are 4GB x 2, ECC enabled, Industrial Grade. The unit runs headless after the initial setup was done to disable all unnecessary services. It occupies just a little over 3.4GB. I have it powered with a linear power supply that can handle 4A (AMB Laboratories σ11 with custom cooling) and it takes a set voltage of 12V. All superfluous (excessive) partitions in BIOS is disabled such as the WIFI card since it is not used. Now... for the sound.

I had both the NUC7I7 and this unit for about nearly 2 weeks before I decided on one and sold the other one off. It was easy to tell from the begging that I liked the iBOX-V1000 much more. It wasn't a SUPER DRAMATIC difference but I would say a 10% gain is quite significant for a product that has the same price and comes in a passively cooled fan-less case. The first thing I noticed was the imaging, detail, sound stage, and void of digital artifacts. The bass was somewhat more articulate and each sonic timbre more real with the same said about the decay. There was more edge to the sound yet it wasn't sharp. It's like re-polishing the corner of table, re-staining it, and putting a new layer of polyurethane over. It was the same sound I had liked or was familiar with but it just sounded more real yet edgy. Digital, yet analog. It only got better in time. At the 2 weeks mark I sold the Intel NUC off. 

 

Having said all this, I could come with some guesses on why it sounded better. I'm not a computer techie, but the improvements could be related to it's low TDP of 15W yet it has faster processing power. It's an APU so it's latency between CPU/GPU processes could be lower since they're both located on the same die. Since the Intel and AMD unit are both small form factor 4 x 4's, that department ends up in a tie. Since the Intel has a wide input range of 12-24 VDC, perhaps a set 12V has less DC-DC conversion going on. The higher frequency RAM meant nothing since I am using 2400MHz sticks so it didn't matter. ECC enabled for the memory does sound a little better (3%) whereas when used on the music server is less profound (1-2%). Lastly I am using the same power supply for them both as I have two of them so it couldn't be because of PSU differences. I even swapped them but they produced the same results time and time again.

 

So all in all, I highly recommend the iBOX-V1000 for those wanting to upgrade from their NUC7i7. The ECC didn't play as huge of a role as I thought but little did I know, every other component was what lead to the rest of the benefits. In regards to components, I advise you to have at least 3A of current for your linear PSU. I found that it sounded better with a 4A supply and the case is probably true when going even higher. I was going to add an MPAudio HPULN (5A) after the Linear PSU I'm using now for extra PSRR and lower output noise to see if it helps with sound (which im sure it does), but in that process I had a different idea. I looked at the unit and hold myself.. "having the endpoint attached to sCLK-EX to replace it's clocks wasn't good enough". I didn't want to improve the clocks of an area that could have been better isolated. For that I felt like the best course of action was to find a way to install both an Ethernet card and USB card into the unit to further isolate the Ethernet and USB controller so that even less jitter could be achieved. This way when the sCLK-EX connects to it, the clock signals are received in optimal fashion since the receiving circuit is now much more quiet. Then comes the discussion on the new USB output's ability creating a more ideal USB signal for the DAC. Even though I have a SOtM tX-USBultra after the endpoint that does some reconditioning to the USB signal, maybe it would do a better job if the signal arrived in better condition than it had before. Or maybe Ethernet entering the endpoint through the isolated card can now be processed more efficiently since there is a direct path to the CPU rather than running it through noisy motherboard components that are all sandwiched in the one area.

 

For this reason I came to the conclusion that despite it sounding better, I am going to go further. For those of you who want the iBOX-V1000, don't let this discourage you. It does sound better, but I think more can be squeezed from an endpoint.

 

I did a search and found that I could purchase a motherboard with the same CPU that also has a set 12V input and all the goodies that was spoken about previously. The unfortunate thing here is that the motherboard is no a longer small form factor 4" x 4". It was bigger, possibly require more power, and having more components on-board for processes that aren't used may create more noise. BUT what it did come with was a PCI-E socket for an add on card. The chassis I found does have a slot for an add on card and passive cooling.

 

This is currently what I have found. I am trying to find a motherboard with the V1605B APU but with 2 PCI-E slots as well as a chassis that can hold both cards while offering passive cooling.

 

Chassis: 

https://www.shop.perfecthometheater.com/FLM-7-Black-Fan-Less-Mini-PC-HTPC-aluminum-chassis-FLM-7-Black.htm?fbclid=IwAR1ZHJMsuS5CrmH1d6hNL7KqxdzyFHd-WT9yz_GsQkDQm2EP3nGqIXDcFLw

 

Motherboards:

[1] https://www.dfi.com/product/index/1404?fbclid=iwar1ww2swkbcrilb5dnpxw_zrsc71t-phnwc-dztnh1yx3smmzc6tk9iu6wi

[2] https://www.scan.co.uk/products/sapphire-ipc-fp5v-1ge-amd-ryzen-v1605b-embedded-cpu-vega-8-graphics-2x-1-gbe-4x-displayport-14-ddr4?fbclid=IwAR19uboDW_ialHLQKCLL30s3KKq99QubFg2aHJjDFDKGFczA9F2DuMzN1h4

 

So there you have it. This endpoint is currently satisfying me but for the future I would like to have a similar one but housed in a larger aluminum case (black) with two JCAT add-on cards for Ethernet and USB. This however would require a larger linear PSU that would have to output at least 5A so I am hoping Farad will build something that can accommodate. 

 

[Energy]

New ATX12VO Spec's

https://www.pcworld.com/article/3518831/how-intels-changing-the-future-of-power-supplies-with-its-atx12vo-spec.amp.html?fbclid=IwAR2iLZt41xB3KZGkXRjt6hVE5oCqqPJrlb9YR58EmUO-Tr-y6Zwk8RrVvdM

 

If they are placing the 3.3V & 5V's on the motherboard instead of having the rails on the power supply..this means more DC-DC converters. It's going to be tough building an audiophile PC unless that section is bypassed.

[Energy]

On 3/25/2020 at 3:30 AM, Gavin1977 said:

Did you manage to find a price for the DFI GH171?

 

Not anymore. I've decided to leave the NUC route and make an endpoint out of the following:

Build:

- HDPlex H3 V3 Fanless Chassis

- X570 Micro-ATX Motherboard

- AMD Ryzen 7 3700X

- Innodisk ECC/Industrial SODIMM DDR4 RAM

- JCAT NET Card FEMTO

- JCAT USB Card FEMTO

- Clocked with sCLK-EX (from SOtM tX-USBultra)

- AudioLinux Operating System (Headless)

Power Supply: (dual regulated)

- Belleson SPHP 12V (14A)

- Belleson SPX 5.0V (7.2A)

- Belleson SPX 3.3V (10.91A)

 

Only this way can I have two add-on cards with the addition of them being directly connected to the CPU.

I just need to find a suitable X570 motherboard first. Since the job of the endpoint isn't as severe as the DSD upsampling music server, I've decided on a Micro-ATX form factor despite it offering less VRM's.

 

For my main system I went with Gigabyte X570 Aorus Master (and Xtreme) because they are the first motherboards in existence to offer TRUE 16 PHASES. What this is is that the PWM controller (voltage regulator) has 16 lanes that connect to the MOSFET's that are used to feed the CPU. Normally in any advertised 16 phase, there is only actually 8 lanes that have doubler MOSFET's on them. Doubler's take one frequency voltage and divides it into two working frequencies. The other method is using two MOSFET's in parallel to achieve these two frequencies required by the CPU. These methods come with problems. A single device outputting two signals does not perform that good and two in parallel sharing one lane comes with it's own problems as well. This motherboard that I am speaking of however has 16 lanes with each an independent MOSFET handling only one job. There is no overworking of a single transistor or two that is bottle-necking one lane and having poor communication with the PWN controller. This improvement leads to higher efficiency, lower voltage tolerance, lower latency, and perfect load balance. To put it short, on paper it should be the best sounding motherboard for audio use. 

 

After the PWN controller and MOSFET, there is an inductor and solid polymer capacitor that filters out the high frequency noise away from the CPU circuit. I plan to increase the capacitance of the solid polymer capacitor. Increasing it cost money but as an audiophile I am not limited to what the manufacturer's BOM was limited to. Increasing the capacitance helps lower the voltage ripple noise and gives the CPU a higher capacity reservoir for pulse currents. I'll then bypass/shunt each of them with ceramic and film/foil capacitors to lower the ESR further so that more high frequency noise can be removed from the system. This further attributes to a lower ripple although most of it is highly dependent on the power supplies's rail.

 

The reason for using different types of capacitors is that their ESR depends on capacitance and frequency.

 

Brief: There are various areas that can be improved on the motherboard. It's just not since they're marketed towards the average consumer hence having to fit tightly to meet the bill of material.

[Energy]

The lower phase ITX could have performed better due to not using doublers and or parallel connections. Often times when there are too many that are not in use, the PWM Controller (voltage regulator) turns off some of them. Which lanes open/close at times could be at random similarly to how CPU cores work. What is available happens at varied intervals. The MOSFET's that are sitting further away from the controller could be the ones that are in use. It may not look like a long distance to us but there is a latency lag when gates open and close. Along the path there are different traces the signal must travel through, each attributing to crosstalk and a potential capacitance created between them. The Aorus I mentioned features the newer server grade PCB weaving for less signal loss. The 6 or 8 layers PCB (depending on model) improves crosstalk since these traces no longer have to inter-lap one another so closely. The 2X copper increases signal fidelity and improves thermal dissipation. In the newer X570 designs there is also an impedance improvement for the traces that go from the CPU to the PCI-E and M.2 slots. This benefits the add-on cards and Intel Optane storage. All that I labeled could be a potential link to why those transient attacks and timing quality didn't meet your expectations. Other reasons not yet listed could be difference in inductor value (higher is better), capacitor value, trace/signal arrangement, or natural noise created by the voltage regulator. There are probably even more that are outside of my area of expertise (I'm an audio mixing/mastering engineer).

 

The feedback you received from the linked post, I would say that I agree with everything the writer says.

First I would recommend starting with a clean linear power supply, then pick out a motherboard that has a capable VRM with the benefits posed above, then everything else from add-on card to CPU to Industrial/ECC RAM are of smaller importance. The same can be said for the endpoint. The only time the CPU is as important as the motherboard or clean power is when you are deciding how that CPU will function. For example, If you need it to do heavy DSD upsampling then the motherboard VRM is vital and the linear PSU has to be able to supply the large amount of current required. These three components affect one another which is why they are of similar importance. If the device you are building is an endpoint then the CPU is less important than the motherboard while clean power is more important than both. Although the unit will perform better with more CPU cores even at a lighter workload, the motherboard is more important because of the VRM's, board layout, possibility of handling ECC RAM,  slots availability for add-on cards, and it's functionality plays a bigger role. Lastly clean power triumphs both. This last mention can be seen by those who have upgraded their power supplies for their endpoints and have found it to be a drastic change. In short, a CPU just processes information, the motherboard should be more important since it is the technical grid but most motherboards are made pretty similarly nowadays whereas clean power supplies are not, especially for audiophile. This is why the priority is set this way.

 

For power I do not recommend SMPS's aka conventional power supplies (AX1600i, SuperFlower). Although they are good, they are only good for the average consumer. Their quality is comparable to high quality Seasonic which have under 20mV (20,000uV) of ripple noise, but compared to an HDPlex 400W (3mV), it is 17mV higher. A Paul Hynes have noise in the single digit uV's (probably). It's a big difference. So when it comes to endpoint power, purchasable brands like Paul Hynes, Sean Jacobs, Farad is probably is good as you'll get. There may be some just as good out there waiting to be discovered, but just without a reputation worthy enough for any of us to take a risk. One thing worth mentioning is although Paul Hynes SR7 can do 12A at 12V, it may still be a little low for a music server that wishes to use some of the higher filters and modulators that is provided on HQPlayer, as doing so will require a CPU with a higher TDP. These TDP numbers also lie since my 105TDP CPU actually does 150W during full throttle so if 14A is cutting it close for me then 12A might be a little under but that heavily depends on what CPU you are using. I remember running Poly-Sinc-XTR-LP with ASDM7EC at DSD256 with an Intel i7-9700K before which is also rated 95W but I never measured how much it consumed before switching over to AMD for ECC compatibility.

 

There are not a lot of build-able linear power supplies other than using Belleson or Sean Jacobs (albeit having to get familiar with current boosters). I went with the more simpler path. For a computer, 3.3V, 5V, and 12V voltages are required. With the Belleson I am able to achieve the following:

- Belleson SPHP 12V (14A)

- Belleson SPX 5.0V (7.2A)

- Belleson SPX 3.3V (10.91A)

 

The reason for SPX is that it has less noise and lower impedance than the SPHP but cannot handle as much current. Since the lower voltage rails are not as important, this drops the price of voltage regulator down while offering better performance (-$20). On a side note, Belleson has not tested their SPHP to perform over 10A, but given their datasheet if you provide it with a 5V headroom, meaning the transformer that supplies voltage to the regulator is at least 5V higher, it can potentially do 14A. This much amps at 12V is 168W. The 3900X I use on my music server draws 150W at full 100% load, but during DSD upsampling it is only around 90W. The idea of having the Belleson be able to handle that high of a current is for the sake of headroom. You never want to run a power supply at full throttle.

 

The standard orientation of the build:

IEC > Transformer > Rectifier Bridge > Filtering Capacitors> Belleson SPHP/SPX > DC Output

 

For my more advanced build: (Balanced, Pi FIlter, Dual Regulation)

IEC > Balanced Transformer > Rectifier Diode > 700,000uF Pi Filter > Belleson SPX > Belleson SPX > DC Output

 

Hopefully that was all understandable.

 

[Energy]

lol thanks for the compliment. I thought I'd use the word to be a little more sarcastic. 😄

 

The NUC7i7 with SR7 you have is already beyond good. I had the ASROCK iBOX-V1000 and although it did sound 7% or so better, it wasn't a significant jump. It mainly came from ECC, faster CPU/GPU, and 12V DC input hence less DC-DC conversion. There's not many places you can jump to right now unless you built your own endpoint as you would similarly to a server. But since I haven't exactly finished the build I cannot say if it sounds any better then the NUC method yet.. It could be worse. I first plan to boot it without any add-on cards to make it an even comparison. Seeing if the faster CPU makes a difference than the lower core count one. From there would I start adding in the JCAT's NET and USB card's one by one. In this hobby that consumes a plethora of money, the only way to make these purchases worthwhile is to confirm that it isn't placebo. The only way to do that is A/B each change one by one. A time consuming process.. 😔

 

I'm building another linear PSU of similar quality for the endpoint so I'll most likely create a public thread on how other people can build themselves one as well. It will be comprised of many different levels and options. Those who want to spend more to get more noise reduction can increase capacitance, inductance (for pi filter), boutique parts (iec, fuse, diodes, transformer, capacitor, connectors, wires, etc), or add another voltage regulator for dual regulation. I will break down the science on why each part is used for each area of the build and why it is superior. I'll also be introducing a new method of grounding that people on here hasn't used yet called Hybrid Grounding along with an infusion of John Swensen Shielding Guideline added to the mix that I call Hybrid Grounding JSSG 360°+. It utilizes a 3-Pin XLR connector as the DC connector. High frequency noise that is picked up by the shield is discharged to the shell of the connector. It's a slightly more advanced method of JSSG but can only be applied with a certain connector. 

 

Last but not least, I will share a personal compendium that details every thought out part of an audiophile system. Even from the minute the power cable leaves the circuit breaker to upon it's arrival at the audio room and how things should be optimally connected. To give some examples...why star grounding copper rods spaced 20FT away is recommended, why 3-phase power circuit breakers are worthwhile, why using a 8-10AWG magnet wire pushed through PTFE tubing and twisted tightly counterclockwise as a powerline cable is better than Romex, how to properly use a ferrite core for audio by picking out the right mix to block high frequency EMI, how to improve a ferrite core by cutting a gap into it to reduce saturation, utilizing MuMetal sheet or iron foil to redirect an isolation transformers magnetic field, how to build a music server and power supply, or which boutique part really matters. The list goes on... I only named like a few out of 100. There will be a table of contents that will explain it all. I plan to put it on a PDF.