Jump to content

Building a DIY Music Server


Nenon

Recommended Posts

On ‎5‎/‎14‎/‎2020 at 10:47 PM, dminches said:

The Taiko Audio Extreme is very expensive but when you start adding some of the top available power supplies for the mobo, EPS and cards, the cost of a DIY server is going to exceed 10k.  That isn't $25k but the gap continues to narrow.  I wouldn't be surprised if Nenon's dual Xeon setup will exceed 15k if he powers it with 2 Sean Jacobs DC4s.  

 

 

I have been working on a C621 intel scalable processor system for about 3 years now.

This is emphatically not a

OAudio Ltd.

 

Share this post


Link to post
Share on other sites

Apologies for the post above, hit return before it was completed.

This is the post as intended.

 

 

 

These costs and the price of this server make me smile. I say this with real insight in the area, I have been developing a C621 Intel scalable processor system >3 years now (actually 7 years if you include the research on other audio servers that lead up to this c621 development).

 

The C621 server development has taken a long time because the decision was taken to research and develop key components from scratch where I wasn't happy with the design / sound quality performance purchased components. These are now fully developed PCB based designs. The process consumed literally 1000s of hours of design time and listening tests. Nothing was out of scope of examination and experimentation, there is real innovations in the build now. Interestingly this includes stuff that Taiko said they could not make work to give the highest level of sound quality performance. Here a particular component is developed and integrated into the server, done correctly it turns out to be exceptionally important to sound quality in c621 systems !

 

So is 29k necessary to obtain a c621 music server capable of top level sound quality, simple answer is no. I can say with confidence it really doesn't require the tick list of expensive unobtanium parts or casework that a tank could be parked on to achieve supreme sound quality. These items might be comforting to potential customers but with different design not needed for ultimate sound quality.

 

Back to the theme of the tread, its really good to see people getting round to using the c621 architecture as a base. Yes the costs are higher than other PC platforms but the they have absolutely magical sound quality potential. More people should consider jumping in.

 

OAudio

OAudio Ltd.

 

Share this post


Link to post
Share on other sites
4 hours ago, Dev said:

 

In our audio application the chipset/PCH is actually doing nothing, if you are using a JCAT net card and USB card on a direct CPU slot. There is no fan to be connected to any of the fan header, there is no sata drives, on board audio can be turned off, etc. So in reality and in our application, the chipset is redundant -

 

 

 

I think this may be a misleading assumption about the role of the PCH.

 

What the PCH does is essential for sound quality and one of the key limiting / enabling factors in the quality of audio achieved by an Audio Server. This is regardless of the route chosen for USB and network connections (eg by On chip PCH or by PCIe expansion cards). 

 

To illustrate the point, I would suggest that pretty much one of the first sub-systems to look at when selecting a motherboard is the clock subsystem of motherboards PCH architecture. In most of the more recent PCH designs for, Intel at least, from the X99 PCH onwards, the clock generation for key the sub systems CPU, PCIe, Sata, CPU to PCH HSSI channels and in some cases USB and network connections are all generated and distributed across the motherboard by the PCH chip. This remains critically important to sound quality even if USB and network are handled by expansion cards.

 

I have highlighted clock subsystem above as one example of why the PCH matters but there are many more benefits for sound quality that PCH designs can deliver. As mentioned earlier at the moment the c621 architecture is in my opinion, by a large margin, the best available right now.

 

OAudio

 

Having said this, research and direct experience here is limited to Intel Chipsets so its not sure if AMD PCHs (X570) behave differently. 

 

 

 

 

 

OAudio Ltd.

 

Share this post


Link to post
Share on other sites
2 hours ago, Downtheline said:

Have you compared the c621 to the c622 for audio quality?

 

 

I looked at specs of the versions of the c620 series at the time the platform was released. I tend to go with the principle that it good not to add things you don't need. The attached table summarises the differences in the family moving from the C621 to C629. The going from c621 to 622 adds 2x10G network and 8x PCIe uplink to the C621 capability, but I don’t need them so C621 was selected to avoid unused PCH complexity. As a side point I have also not had good SQ experience with 10G network cards for various reasons so this was another reason for going with the c621.

 

Having said all this I haven't physically tried the c622 so it may sound better but I am more than happy with the c621 performance.

C620 family.PNG

OAudio Ltd.

 

Share this post


Link to post
Share on other sites
2 hours ago, lmitche said:

Here is a two year old article discussing the difference between the Intel and AMD clock implementations.

 

https://www.anandtech.com/show/12678/a-timely-discovery-examining-amd-2nd-gen-ryzen-results

 

It is clear that the OS and BIOS configs have a big role in determining which clock is used. It looks like the industry is moving to TSC clocks onboard the processor chip.

 

My AMD motherboard running Audiolinux shows the TSC clock is used, which is a clock onboard the Ryzen 7 2700 processor chip, so the PCH is not used. In my case the PCH is disabled.

 

#cat /sys/devices/system/clocksource/clocksource0/current_clocksource
tsc

 

 

 

 

I'm reading up in the background to try to come up to speed a bit with the AMD stuff :-)

 

HPET discussed in the the linked article and TSC are event timers (circa 1ms upwards) that are referenced to a physical oscillator (either PCH attached for Intel or internal to the CPU for AMD). So I get that this is CPU attached in your system.

 

If I understand, I guess your saying that in your system PCIe attached devices (USB and Network) are direct channels to the CPU, you have no storage attached via PCH Sata as well and the TSC timer is integrated onto the processor. Very neat :-)

 

Is there a mode that can explicitly shut down the PCH on this type of AMD system ? I'm just thinking that the PCH will still be live and there will be some traffic between it and the CPU, just not much in the case of you set up.

 

 

 

   

OAudio Ltd.

 

Share this post


Link to post
Share on other sites
3 hours ago, Dev said:

 

can you provide more details on what c621 board are you using and what cpu, chassis, cooling, power supply, etc ?

 

I may start a separate thread.

Neons build is following a different and interesting route so I think it would get confusing and unfair to get into detail here.

OAudio Ltd.

 

Share this post


Link to post
Share on other sites
On 6/21/2020 at 8:37 PM, Nenon said:

You are absolutely right. And the Extreme is a great value when you consider all the parts inside, the time you would save, the support you get, etc.

 

I am not so sure about this. I actually think if anything was compromised on the Extreme, it might have been the power supply. But I don't know much about the design, and it would not be fair to make any conclusions. So far I have not seen any compromises on the Extreme, and maybe if I learn more its power supply I would change my mind :). We know Emile is full of surprises!

But I can comment on what I see and if I was to do some educated guesses and speculations, here is what I think. I'll break this down in 3 parts - capacitors bank, transformer(s), and number of rails.

1. Capacitors bank.

What I see is 14 electrolytic Mundorf Mlytic AG capacitors (those are the big capacitance caps) and 3 small value film capacitors (2 Duelund and one Mundorf in a special tankwood stand). Let's focus on the electrolytics as they provide the high capacitance. Those Mundorf Mlytic AG go from 1000 uF to 47000 uF. The most commonly used values in DC LPS are 10,000 uF, 22,000 uF, 33,000 uF, and 47,000 uF. They are typically 30 mm or 35 mm in diameter. I have seen Sean Jacobs using different values in his power supply. He uses four of these per DC rail in his DC4 (and in some DC3's). I don't think he uses the largest values, and I believe this is done for a reason (not related to cost).

A perfect DIY server has 6 rails - 3 rails for the ATX, 1 rail for the EPS, 1 rail for the Network card, and 1 rail for the digital output card (i.e. USB). So, going with 6 x DC4 rails, you will end up with 24 of those Mundorf Mlytic AG capacitors. That's actually a lot more than the Extreme (14). I am not considering the value of the capacitors here. I think Sean is using 10,000 uF by default, but I also think if you need more capacitance he would be able to accommodate it. Whether the end result would be better or worse, I don't really know. I am strictly comparing the capacitance in a vacuum. 

 

2. Transformer(s)

Taiko Extreme uses one 400VA transformer. That's probably not the best transformer for the Extreme. I do believe the transformer is a weak point here.  I power up my dual Xeon server with two 400VA transformers that I believe are of better quality, but I don't want to judge by the cover. Taiko has the transformer geniously installed in a tankwood enclosure - that's brilliant! 

I also always find the transformer to reduce the sound quality when installed in the same chassis as the motherboard. But the Extreme has a thick copper panel between the two. I have tried multiple different panels - they helped but did not completely fix the problem - but I have never done such thick copper panel or the many holes on the Extreme designed to help with EMI/RFI. Let's assume Emile has nailed this down, and that has no negative impact in the Extreme. 

But what if you were to go with 6 rails of DC4? Per Sean's web site, the DC4 is limited to 3 rails: "If you require more than 3 outputs, then you will need to order multiple DC4 units". So, that means you actually need to get two DC4 power supplies for 6 rails. 

That is 2 x 600VA transformers! Again, much (3 times) more than the Extreme. You don't get the tankwood enclosure, but the DC4 has some vibration damping... probably not as sophisticated as on the Extreme though. 

 

3. Number of rails

From our past experiments, we know that the more isolated our server components are with different linear power supply rails, the better. A separate rail for the EPS makes a big difference. A separate rail for the USB makes even bigger difference. And so on. That's why my perfect build has 6 separate DC rails. That has always been better in all my tests. The problem with so many DC rails is heat dissipation. 

I have no idea what Taiko Extreme does for power supply, but I honestly see one rail only. If the Extreme was using many rails with linear regulators, I would expect to see those regulators exposed to the heatsink with adequate space between them, to provide proper cooling as they need to dissipate quite some heat. I don't see anything like that there. If I was to speculate big time, the Extreme is using one rail and the everything is handled by switching mode regulators. Again, Emile is full of surprises, so is it possible that he has discovered something that works better than multiple rails of LPS? Why not... I would not be surprised at all. 

 

All I wanted to say is that although that bank of capacitors looks very impressive, there are DIY options than can give you much more capacitance. But the important part here is the end result. Taiko Extreme has set a new bar and our DIY experiments are not even close at that point. We all need to get back to work :). 

@Nenon

 

A few thoughts, its possible you may be looking at something a little different then you are expecting.

 

Caps.

 

The M-Lytics will be 47k uf. The server is described elsewhere as having approximately 0.7f in total. Maths says approx 47k uf per cap give or take. They can be difficult caps to apply in such large size for servers, personally I have found they slow musical response and its difficult to keep life and air in the music using them. Credit and respect to Taiko if they have found a way to apply them without these issues. 

 

Looking at the thermal design and power supply.  Whilst the case has heatsinks that are proberbly ~ 0.5k/w dissipation on each side, there is no dissipation acessable directly from the power section of the design. It may be that heat conduction is through the base of the case to the heat sinks but this may not be not very efficient. If the at least the main rails were powered  by multiple linear supplies dissipated heat would be 40 to 60 watts + or - from linear supplies. It' clear the CPUs are well catered for by the side finned heat sink but looking at the thermal approach my money is not on linear supplies. Based on the thermal packaging and layout of the visible supply elements, I would go out on a limb and guess that there could be buck convertors under the caps, even a DC ATX - which SGM servers have used previously. If you know how to modify them well and apply them optimally their sound quality can be much better than people generally realise so I would not regard this as a limitation if they are used.

 

Transformer.

 

If the thermals do point at bucks for the main high power rails then 400va would be a good size for the power requirements of the server. The choice would be helped by the CLC input filter which will increase the power factor of the supply easing the life of the transformer. I find the sweet spot for server transformer sizing a VA of 2 to 3.5 times the stable VA total draw of the server (consumed + disipated power). Go large and sound becomes slow and stolid.

 

Multiple transformers need careful thought and design if the key aim is to maximise sound quality. The issue is that the power factors of linear regulators that supply key server rails tend to be  low leading to high peak currents in transformer secondaries, particularly when using very large capacitances. Where dual transformers are used it can lead to each transformer having a different distortion caracteristics under their respective low power factor loadings which in turn can apear as differential noise between the rails which are supplied by the different transformers. It took a long time and a lot of work to pin this down. An example sound quality impact might be if you have ever been suprised that you get a better sound powering say a clock from the servers main trafo rather than its own dedicated "clean" transformer, there could be many possible causes but look carfully in this direction first. 

 

OAudio.

 

 

OAudio Ltd.

 

Share this post


Link to post
Share on other sites
On 10/14/2020 at 6:58 PM, Superdad said:

(probably for its 3.3V input, small footprint, programmability,

 

and fancy gold color)

 

:)

 

On 10/14/2020 at 6:58 PM, Superdad said:

For comparison the $10 25MHz Crystek 575s we use are running about -108 to -112. Compensating for the 2.5 octave difference that -110dBc/Hz average equates to about -119 if that 25MHz Crystek was a 10MHz clock.

So why pay $150 for just unneeded long-term frequency stability. For these applications it does not matter at all if the frequency drifts a tiny bit over time!  

 

Absolutely.

 

 

OAudio Ltd.

 

Share this post


Link to post
Share on other sites
1 hour ago, Nenon said:

When you get a LPS from Sean Jacobs or Paul Hynes, you know it comes with many years of experience and expertise.

 

I think it would be good to see measured performance specifications provided as well.  

 

As a general observation there are some large ticket LPSs out there for audio systems but few vendors provide measured specs for products. At the price levels we are operating at, it's not unreasonable to be looking for similar measured performance to lab class LPSs from companies like R&S Tektronics and Agilent.

 

So ball park or better than:

Sub 50 uSec 60% load transient recovery times.

Sub 2mv rms ripple and noise.

Sub 2mv rail sag from 10 to 100% of rated load.

(For 6 to 10 amp rated supplies)

 

I know from developing supplies there are many layers to a good design but my experiance is that these measurements do correlate with SQ and could help better inform people's decisions.

 

 

 

 

 

 

 

OAudio Ltd.

 

Share this post


Link to post
Share on other sites
15 hours ago, One and a half said:

For ATX spec, the challenge for linear supplies is a daunting one.

 

Defiantly agree, but has been addressed. 

 

For the project I am working on reliability and flexibility are two of the key requirements, so this challenge has had to be addressed. I think the result (described below) is possibly one of the most flexible server power system I have come across for high end audio servers. The system is certainly not a set of discrete power supplies, hot wiring ATX control wiring  and crossing fingers for the "Kerr-thunk" as power is applied is too exciting / potentially expensive :(

 

 

15 hours ago, One and a half said:

For (E)ATX boards, it's quite a different kettle of fish. There's 5V, 3.3V notably the 3.3V with 100mV max sag and lead compensation. Even the Power OK signal for a linear supply has specific timing so that the mobo understands what the PSU is up to.  Under short circuits or overloads, the power is limited to 240VA output for any output.

 

Intel ATX Power supply guidelines attached to this post refers. 

 

For ATX spec, the challenge for linear supplies is a daunting one (to an amateur PSU designer like me). 

atxps09_atx_pc_pow_supply.pdf 112.89 kB · 3 downloads

 

 

All very important points that have to be addressed. The solution I have built works as follows:

 

The system is fully modular, so that for my project I am able to reconfigure power arrangements, safely adding and removing LPSs modules / rails as needed. Critically I require real robustness and reliability from the system, so it is designed and implemented to have full control of LPS rail sequencing, it does real time monitoring of all LPS output rails, monitors LPS module temperatures (there can be significant heat dissipation levels on some supplies), and has system lock-out levels with error reporting if something exceeds tolerances I have set. 

 

The power system was in part designed as a development tool, to support safe working through a broad range of power configurations to develop a music server. To hit this requirement the power system can be be configured with 1 to 7 of the high current high LPSs modules / rails, all managed as above. In addition the system was designed to work along side other supplies types to allow hybrid power combinations (eg mixes of linear / buck / SMPS), so it works with one or more DC-ATXs and or SMPS ATXs if needed. In a hybrid setup the power system operates either as the master power controller (with up to 7 LPS modules and DC-ATXs), or as a slave to a DC-ATX / SMPS ATX (but still managing between 1 to 7 LPSs as required).

 

As you will see there is a long list of possible configurations that can be safely set up and tried and a very significant amount of time has been spend working through what delivers SQ whilst developing the server.

 

OAudio

 

 

OAudio Ltd.

 

Share this post


Link to post
Share on other sites

 

19 hours ago, lmitche said:

Are these targets or actual measurements?  

 

The first 2 sets (copied below again) are my personally thoughts on what I would look for. The first case 1) for me would get the job of powering a DC-ATX or a single music server rail done, whilst the second case 2) would be for a more expensive class of device.  Transient response, ripple and sag under load are quite basic parameters but I think they do have some correlation with SQ, and so are useful. The last set of figures 3) below are actual measurements taken from the 12v, 10 amp LPS module of the system I have built.

 

1) - to get the job done.  

 

"Start at < 200 uSec for 5 amp transient rail voltage recovery with a critically damped response and <20mV over / undershoot. This is is typical for low to medium cost bench linear PSU." 

 

2) - at a higher price point  - looking for comparable performance to descent lab level LPS's.

 

"2Sub 50 uSec 60% load transient recovery times.

Sub 2mV rms ripple and noise.   [edit ** actually, sub 1mv would be more realistic.]

Sub 2mV rail sag from 10 to 100% of rated load." [edit ** again a little low ball, sub 1mv more realistic.]

 

Above both for LPS's in the 4 to 10 amps continuous current range.

 

 

 

3) Taken from the 12v 10 amp LPS rail in server power system I developed. 

 

"Load rating: 10amp continuous, 14 amp peak sustained draw. (Transient (mSec) current rating is far higher but its not been measured beyond 25 amps. 

Transient recovery*:   45 uSec for recovery to rail set voltage. Transient output voltage over / undershoot < 1.5mV (see note below).

Ripple and noise :  Total < 350uV pk-pk measured at 8 amps load.

Output Voltage sag:    1-10 amp sag <500uV

 

*Transients measurements are for a 1 khz square load waveform alternating between 3 and 8 amp load currents. The scope used for measurement is in high res mode....."

 

 

Just my thoughts above, if anyone else wants to chip in feel free. 

OAudio Ltd.

 

Share this post


Link to post
Share on other sites
13 hours ago, sandyk said:

Larry

Some smaller companies will not be able to afford the sophisticated gear needed to do this. 

As an illustration, Uptone took a while to be able to afford the expense of designing and using equipment like that, as it all comes from the profits which permits a company to invest in that type of gear after they become financially stable after recouping their initial investment. 

 

Regards

Alex

 

Alex hi,

 

You may have misconception about how complex / expensive these server PSU measurements are to perform.

 

Server LPSs are not ultra low noise units and the measurements we are discussing are simple, so they are straight forward to make and do not need sophisticated or expensive equipment. An oscilloscope and an electronic load for say £1-2k in total sourced from good manufacturers will do this. Its worth remembering as well that the same pieces of equipment are the tools required to develop, test and produce supplies, so they should already be on hand.

 

I suspect in most cases it's a choice (or oversight) not to publish specs rather than a problem accessing equipment.

 

 

OAudio Ltd.

 

Share this post


Link to post
Share on other sites
14 hours ago, dminches said:

The market is wide open for someone else to build a DC-ATX converter.

 

 

That is an interesting comment. Do you mean

 

A DC-ATX using switched buck converters eg another HDPLEX type device.

 

or 

 

A true fully linear DC ATX system that plugs into the motherboard and is fully controlled, managed and monitored as an ATX supply ?

 

 

OAudio Ltd.

 

Share this post


Link to post
Share on other sites
21 minutes ago, MarcelNL said:

 

 

EDIT:also, does anyone have thoughts on the amount of RAM? enough is enough and more adds noise or is it never enough? (I came from 64Gb now 8, which is plenty for Daphile, it shows I have 7Gb free memory)

 

 

 

Hi MarcelNL,

 

The most important thing when deciding how much memory to install is how many DIMMs are needed to have one DIMM per CPU memory controller / channel. For instance on a c621 motherboard there are usually 6 memory controllers / channels so 6 DIMMs are needed. For an X99 / X299 there are usually four controllers / channels so 4 DIMMs. Etc.

 

Best to look at your motherboard manual to work out how to populate the physical slots so as to use all available controllers / channels (physical DIMM sockets on motherboards may not nessasarily be mapped one to one to a cpu controller / channels).

 

Otherwise within reason keep the capacity of individual DIMMs low 2 or 4 GB per stick seems OK.

 

OAudio.

OAudio Ltd.

 

Share this post


Link to post
Share on other sites

Try 1 dimm per channel and then 2 just to check out any sound quality change.

 

Often the extra dimm on a channel dosent add anything or makes sound a little worse, but that might not be the case with your board 😉

 

 

OAudio Ltd.

 

Share this post


Link to post
Share on other sites

+1 here for the 665p Intel drives. Their SQ is usage dependant, which isn't really a suprise, but they are capable supporting very good sq when used for media storage. 

OAudio Ltd.

 

Share this post


Link to post
Share on other sites
1 hour ago, ASRMichael said:

Hi, are you tweaking the Extreme? 

I have servers in the late stages of development, there are single and dual socket c621 varients. SQ is comparable with the Taiko.

 

Of the single and dual socket setups the dual socket is definatly the most felxible. As ray-dude says, it's satifying to hear 

SQ benefits when data paths and software processes and priorities are well setup.

OAudio Ltd.

 

Share this post


Link to post
Share on other sites
1 hour ago, ray-dude said:

I suspect there will be a part 6 to my Extreme review at some point (one year in type of thing), once the tweaks and updates accumulate a bit more.

It's a great review serise, looking forwards to part 6 !

OAudio Ltd.

 

Share this post


Link to post
Share on other sites
2 hours ago, MarcelNL said:

I wonder if anyone here has experimented with power supplies?

I am quite sure that any switching supply with lower ripple and igher power rating is cosmetic dabbling, I'm not yet so concinced that ultra low ripple is more important than a stable voltage.

 

Making a whopping 12V C-L-C PSU should prove doable, def more doable than a high current lineair well regulated low noise PSU. Ideas, pointers, links are welcome?

 

 

DC ATX supplies can put in good performances but PC's are current driven devices and at the current levels required DC ATX and other switched / buck supplies will only take you so far.

 

Linear is perhaps the best route but it's a significant task developing them for ATX supplies as an application. CLC is attractive but can occupy a lot of space and needs to be carefully designed but can be done.

 

We have gone for multiple linears with a list of special features in the supplies. Performance even at high(ish) currents is good and it shows in SQ achived.

 

If you can find specs or if you build try to target the measured performance recommended here (see link below). Measured PSU performance is not everything but this is one of the rare areas :D in hifi where good figures do come through into SQ.

 

(As an aside the current figures for our supply have been re-rated. From 10amp to 15 amp contunuous and 22 amps sustained peak per module / rail. These levels are not really needed but its nice to know that the headroom is there.)

 

  https://audiophilestyle.com/forums/topic/58164-building-a-diy-music-server/?do=findComment&comment=1086986

 

OAudio Ltd.

 

Share this post


Link to post
Share on other sites
4 minutes ago, The Computer Audiophile said:

Hi @OAudio, I just want to make sure you are still just a DIY'er and don't have a commercial offering yet?

Hi Chris,

 

No commercial offerings yet but the situation could change moving into next year so I will pm you for guidance.

 

 

 

 

OAudio Ltd.

 

Share this post


Link to post
Share on other sites
4 hours ago, Nenon said:


Active rectifiers are not better or worse in my experience, but they sound different, so just another way to ‘voice’ your power supply. They however have a benefit of lower voltage drop and less heat dissipation which might have bigger impact than the rectifier design itself.  

 

Active rectifiers can provide real up tick in SQ, but to be more confident of them doing this, they need to be designed into a supply from the outset.

 

When swopped into existing circuits, particually higher current LPSs which were designed with diode bridges originally, the impact on performance and sound quality is variable. 

 

For higher current capacity supplies at least, keep in mind that dropping in an active rectifier can generate a really very significant jump in peak ripple current in the transformer, rectifier and filter components. The 'on' resistance (Ron) of an active recifier can easily be less than a tenth of the ROn value of the diode bridge being replaced. The components of the supply (transformer / filter caps in particular) may or may not deal well with the change in peak ripple current and this can pull SQ across the dial. Can be better, worse or different depending on the situation. I think its still worth experimentation swoping out diodes despite this.

 

Active rectifiers can generate a excellent sound quality up tick. But to be more confident of getting to this, the LPS has to be designed to deal with and exploit the low Ron of the active rectifier (which can easily be just 4 milli ohms). Then there are the other benifits such as the low switching noise of active rectifiers can also come into play.

 

OAudio.

 

 

 

 

OAudio Ltd.

 

Share this post


Link to post
Share on other sites


×
×
  • Create New...