DIY Project High Performance Audio PC with high quality wiring

[Nenon]

Let me be the first one to comment. Happy to see more builds documented, and this one looks promising already. Please see my comments below.

 

8 hours ago, StreamFidelity said:

1.4 Corsair DOMINATOR® PLATINUM 16 GB (2 x 8 GB) DDR4 DRAM 2400 MHz C10 memory kit

 

Due to their robustness and low latencies, these 16GB RAM are also well-known.

Have you considered using Apacer RAM? I think you will like it.

I am pretty much done trying consumer memory. No matter what speed or latency they have, the Apacer always sounded better in my system. My latest discovery was that you can use ECC RAM on those LGA 1151 motherboards. It works as non-ECC RAM but sounds a little better. 

 

8 hours ago, StreamFidelity said:

1.6 JCAT USB Card Femto

 

Experience shows that a dedicated USB card with an external power supply is worthwhile. At least when the DAC is controlled via USB. Other cards are also conceivable as an option. For example from Pink Faun.

You may want to hold off to see Marcin's new reference USB card. 

 

8 hours ago, StreamFidelity said:

1.8 HDPLEX 400W DC-ATX

 

An ATX mainboard also requires an ATX power supply. The HDPLEX 400W DC-ATX converter has proven itself very well here.

If your budget allows, try the HDPLEX 800W DC-ATX. It fits nicely in the chassis with the ASUS ROG motherboard:

But you will have to drill holes on the bottom of the case. And yes, I will have a super short ATX cable, that's why it's lined up like this . 

 

8 hours ago, StreamFidelity said:

The DC power sockets still to be selected on the HDPLEX fanless chassis should, however, be flexible in use so that any power supplier can be connected.

I ordered some Jaeger connectors and a thick carbon fiber plate. Planning to replace the stock small power connector plate that comes with the case with a custom carbon fiber plate and two Jaeger connectors on it. I will post more about it when I make it. I think it will look really good. And those connectors are top notch. 

 

8 hours ago, StreamFidelity said:

1.11 Minimization of resonances

 

Sensitive electronic devices particularly benefit from stable, resonance-free device feet. We will experiment with different BF Magic Spacers from fis Audio. Balls made of glass, precious metal or other materials are placed on the discs and deflect vibrations.

This is interesting. I may ping you on PM for more details. 

[Nenon]

1 hour ago, adamaley said:

Does anyone have an idea when the HDPlex 800W DC-ATX will be back in stock from HDPlex? Or is anyone looking to offload one they're no longer using? 

It was in stock yesterday. Email them, maybe they have some. 

 

24 minutes ago, rafa said:

Because you should slow them down, release timings and then RAM sounds better.

 

Faster = more technical sound

Not from my experience.

The 2666MHz Apacer RAM running at 2666MHz sounds better than the 2400MHz Apacer running at 2400MHz or any slower speed. 

Also, the 2400MHz Apacer RAM sounded better than my 3200MHz G.Skill. The most important property for RAM to sound good is its quality. It's not the only factor but the most important one. 

It also matters how you tweak your server. If you are using a low-powered CPU, slowing down the RAM helps. I am not talking about those types of servers. Gave up on them a while ago but I do remember that slowing down the RAM helped there. My comment was about @StreamFidelity build, which is not aimed at low powered CPUs. With those servers, typically (given equal quality) the higher the speed, the better it sounds. And by better, I don't mean "technical sound". 

[Nenon]

1 hour ago, dminches said:

 

Did Larry tell you this?  He owes me one since the one I bought 6 weeks ago was defective.

 

I was checking something on his web site yesterday and saw them showing as in stock. 

 

1 hour ago, auricgoldfinger said:

Marcin found this 2400MHz UDIMM to sound the best.

 

http://jplay.eu/forum/index.php?/topic/2786-ram-adventures-extremely-interesting-results/page-2#entry49239

I see other people commented on this already. Marcin's post was about the 2666MHz RAM. That was the "new" RAM he referred to. 

 

29 minutes ago, dminches said:

What is the downside or lost opportunity in using ECC ram if the mobo uses it like non-ECC ram?  If one is using ECC ram shouldn’t they try to find a mobo that supports ECC?

I don't really know but the ECC RAM sounded better as non-ECC on the last build I posted about. Perhaps it's not because of the ECC bit but the quality of the memory. I compared a 2666MHz Non-ECC Apacer with the same version of ECC 2666MHz Apacer running as non-ECC. I liked the ECC RAM running as non-ECC better. 

 

There is another member on this forum performing the exact same tests (2666MHz Non-ECC Apacer vs. 2666MHz ECC Apacer running as non-ECC) but may take him some time to report. 

[Nenon]

On 2/22/2020 at 4:30 AM, rafa said:

I tested 6 different DDR3 RAM modules and agree that it's quality most important.

 

But about timings I have opposed opinion. Maybe that depends on definition, what is "better"  

 

DDR3 RAM?!? You, should have started with that. We have been talking about different memory type then - completely different ways of optimizing the two. 

[Nenon]

3 hours ago, motberg said:

A friend's research (using a bunch of MPAudio LT3045's and other regulators ) concluded that in his build the 5V was most important to the sound quality

This is interesting. In my tests, the EPS connector is the most important to the sound quality. Can you please share what motherboard and CPU your friend used in his testing? I remember that on low powered motherboards the 5V made bigger difference and benefitted by oversizing the power supply a little. The reason for that might be that the oscillators might be using that rail, but that is just a guess. 

 

3 hours ago, motberg said:

just looking at Ghent's prices as a base reference... I guess the fis stuff is going to be somewhat expensive 😁

The Neotech wire is $3 per foot, and that's the best conductor Ghent offers. The Mundorf silver/gold wire I use in my DIY DC cables is $33 per foot. 

 

Although this thread is intended to focus on commercial cables at some point, I would like to point out that I have not heard a better DC cable than the DIY recipe I have shared with everyone yet. I do encourage people who are looking at the best of the best to give it a try. The best thing (for some) is that this is a DIY cable that you can do on your own, which makes it a great value for the money. The bad thing is that it's a DIY cable, which means not everyone can make them.

 

@Blackmorec wrote an awesome review about a week ago and told me today they got even better:

 

@Dev made one of them too and liked them better than his reference AudioSensibility solid core silver cable.

 

Another member pinged me a couple of days ago and told me that the DC cable made much bigger difference than replacing the regular RAM with Apacer RAM. And for those of you who have tried the Apacer RAM, you know than means a lot. 

 

I am following this thread with pleasure and waiting to see what fis would come up with. Anything that sounds better is always welcome. The more good products we have, the better!

 

And just for completeness, you can find the DIY DC cable recipe by following the link in that post:

 

[Nenon]

14 minutes ago, motberg said:

Yep - he was using a small server MB, not sure the model but it was one of the small Supermicro server boards....

 

 

4 minutes ago, dminches said:

What makes a motherboard "low powered?" 

 

I was referring to some tests with the SuperMicro A2SAV-L motherboard, similar to what Innuos uses. 

[Nenon]

Why don't you rotate the HDPlex DC to DC convertor 180 degrees and move it to the left where the AC to DC adaptor is marked to be? Something like this (although this is the 800W DC to DC):

 

 

This would keep all cables shorter. The downside is that you may need new holes, but I think it's totally worth the effort. 

 

[Nenon]

On 3/5/2020 at 8:13 AM, StreamFidelity said:

All cables used are "heard by direction" and are also assembled in this way.

If you determine the "right" direction for the plus, do you reverse the direction on the minus cable?

 

On 3/5/2020 at 8:13 AM, StreamFidelity said:

The Molex plugs with the inserted pin contacts are basically too shaky for him. He is planning a wobble-free connection.

The Molex branded housing is actually better than some third party ones. I was considering filling in with epoxy, but I haven't found an easy way to do that. It would be interesting to know what your approach would be.

[Nenon]

14 hours ago, Triplefun said:

the AMD Ryzen 9 3950x with 16 cores (32 threads) at 3.5 ghz and only 105w tdp - runs a LOT cooler than the i9 9900k

 

Is that true? Can you share more info about it? I am not that interested in reducing the CPU voltage... but even if it is running as hot as the i9 9900K, not "a LOT cooler" that would be really good. 

My recollection was that @romaz had to disable the turbo, significantly reduce the CPU frequency, and possibly reduce the voltage on the AMD Ryzen 9 3950x in a Streacom FC9 chassis to keep it relatively cool. At that point I would stick to the AMD Ryzen 7 3700x actually. 

But I would be interested to know where you got that datapoint from. 

[Nenon]

2 minutes ago, sandyk said:

 

 I used to think the same, but Blacmorec's post 83 makes sense, especially if using gold plated Molex pins with their leads soldered;

 

I agree as well. Gold plated Molex pins first crimped and then soldered with WBT silver solder is what I do. 

[Nenon]

This is turning out really nice. 

 

On 4/23/2020 at 8:02 AM, StreamFidelity said:

It was very difficult to attach the back wall with the three PCI Express cards.

YES!!! I guess I am not the only the one thinking that removable back plate is the worst part of this otherwise excellent H5 case. 

[Nenon]

1 hour ago, StreamFidelity said:

I sawed off the metal web above on the back plate, so the PCIe cards can be replaced quite easily. 😁

 

Good idea! 

[Nenon]

You can see it if you zoom in the picture above.

 

[Nenon]

@StreamFidelity - I've had a few people asking me which DC cables make the biggest difference. 

 

I think you have 5 main cables here:

1. HDPlex input

2. ATX cable

3. EPS cable

4. JCAT USB cable

5. JCAT NIC cable

 

If you have a chance, can you give us your subjective opinion which cable is most important? For example:

- If I can get only get one premium cable, which one would it be?

- If I can get two cables, which ones would you pick? 

- If I can get three cables? 

- Which cable would you leave in stock form if I could get four cables?

 

In my system, the biggest difference makes the cable that powers the USB card, followed by the EPS cable (not going through the HDPlex). That's confirmed by listening tests. But I have not compared the ATX cable and the HDPlex cable, and it's hard to rank the impact of those two. 

 

Given the title of this thread and the focus on high quality wiring, I think it makes sense to try to answer some of these questions. Your friend who makes the cables would also benefit from that as it would probably be a frequently asked question. 

 

[Nenon]

Sean's latest "10A" LPS is capable of delivering 30-45A. The limiting factor is the heat dissipation of the case. It's not listed on his web site, because he is still perfecting it. It's based on his DC4 design and it's a cost no object design with dual regulation and all the custom stuff he has developed for the DC4. 

I have been a beta tester for a while. Waiting to receive his latest prototype in a HDPlex H5 chassis to see how well the heatsinks on the H5 would do in this application. It may need to go in a case with better heatsinks at the end. The Asus SAGE / dual Xeon box I built is perfect for testing those types of power supplies - highly resolving and drawing a lot of current if unleashed. 

 

Sean also offers a DC3 based 10A LPS, which he does not advertise actively, but I know quite a few people are using it. It can also easily handle peaks over 10A.  

 

Paul Hynes Design SR7 is a cheaper alternative that is rated at: 12vdc @ 12A continuous, 40A transient. By cheaper, I don't mean worse. I have not had a chance to test this standard SR7 yet, but the feedback has been good.

 

Keces P8 is a lower end LPS. I have built some computers with it. It does a good job for its price. But there is a significant difference between the P8 and Sean's DC4 in both - price and performance. 

 

Those are the 3 high current power supplies currently on the market that I would consider. They represent 3 different price points. It really depends on your budget and your system. They are all good options. A couple other well-respected companies are also working on new high current products. It's good to have variety!

[Nenon]

3 minutes ago, dminches said:

Where do the HDPlex 400W and 300w (I guess this is new) fall within the units you mentioned?

 

Somewhere with the Keces P8 IMO. Perhaps someone has done a direct comparison and can tell more. I've never had them at the same time. 

[Nenon]

25 minutes ago, Exocer said:

Do you still feel the need to power PCI-E expansion cards (JCAT Femto NET/USB) with dedicated LPSU if the PC is already powered by an lpsu at the Keces/HDPlex level or better?

 

Let me answer this, because it's an easy question. ABSOLUTELY!

[Nenon]

16 hours ago, StreamFidelity said:

Nenon has published his prices here:

Just to clarify - these are not my prices. I have no intentions to build computers for other people anymore. I put this post together to give a general idea about the price of a DIY server... done as DIY, not by me. 

[Nenon]

5 hours ago, StreamFidelity said:

I became aware of Solarflare through these report:

Thanks a lot for this. At first there were unfortunately some failures. The Solarflaire Series 7000 could not be operated in the fanless audio PC. These cards just get too hot and then switch off. But what I had heard so far in terms of clarity and transparency left me with no rest. And so I ordered a brand new card from the 8000 series: XILINX Solarflare Flareon Ultra SFN8522 - network adapter - PCIe 3.1 - 10 Gigabit SFP +.

 

This card has an even larger heat sink and is generally very large.

 

 

 

 

The back is also well equipped.

 

 

 

Ventilation

 

In the accompanying note it was pointed out that a fan with at least 200 rpm is required for cooling. That made me worry, of course, because there is no such thing in the fanless fis Audio PC. On the other hand, 200 rpm is very low. A 24h burn-in test has so far shown no abnormalities. Due to the principle of convection (chimney effect), warm air escapes upwards through the ventilation slots and draws cold air after it. That seems to be enough.

 

Installed it looks like this (the map on the right). The air supply is important. The left empty slot will be equipped with another card in due course. So there follows another report. 😉

 

 

 

Specifications

 

All eight PCIe lanes are used for data connection to the CPU, so a slot with a direct connection to the CPU is recommended. The latency is reportedly in the very low range of less than 1μsec.

 

According to hearsay, Solartech network cards are also used for high-frequency trading on the stock exchange. 😁

 

The product letter can be downloaded from XILINX.
 

Drivers and manuals

 

The drivers are available here. The Windows Server drivers also work perfectly in Windows 10 Pro. The manual is extensive as usual with around 300 pages. There are even recommendations for low latencies.

 

Of course there was try & error

 

- PCI Express Lane Configuration
Ensure the adapter is in an x8 slot (PCIe Gen 2.0 or PCIe Gen 3.x) or x16 slot.
Bad trap. At the beginning I got a slot with x4. Of course it didn't work.

 

- Transceiver
My previous SF transceiver module 10GBASE-LR SFP + 1310nm 10km did not work. But the Startech 1000BASE-EX SFP -SM LC-40 KM transceiver. Those are the pitfalls with fiber optics.

 

- BIOS
The Solartech card enables some settings in the BIOS. I got this when my audio PC no longer booted after being switched off and went into Power Safe mode. That's always stupid, because then I can no longer work remotely, but instead had to connect a screen with a mouse and keyboard. Deactivating the UEFI boot via Ethernet solved the problem.

 

My setup

 

 

The picture shows the calm flow of data. No nervous fidgeting like with unconfigured onboard network adapters.

 

 

Sound description

 

The sound is just addicting. A sensational speed. Flamenco or drums are a real delight. This clarity and exactness of the tones are for me the best so far that I have heard on my system. The separation of instruments is outstanding, combined with excellent spatiality. And the whole thing garnished with a beautiful flow of music without any sharpness. Despite the high resolution, voices have a nice melt and the bass is deep and precise.

 

Thank you for the report. Solarflare NICs are widely used in the financial sector where low latency is a key. I've used thousands of them and always had it on my list to try one on my server (i.e. the 8000 series in particular). 

The only reasons I haven't tried is because:

a. It's hard to power externally; and

b. Those tweaked Buffalo switches sound so good! I would like to enjoy them as much as possible before I start a hunt for a 10G switch. And by the way it is worth trying AOC cables here... 

 

So far, pretty much everything that works in the financial market networks I have designed has turned out to work really well in my audio system - multiprocessor servers, NVME storage, low latency tweaked kernel, etc. etc. I should start paying more attention to this. So maybe I would take a 10G Arista switch or a Cisco Nexus 3524 (because of it's low latency - 190/250 ns) and try to modify it. 

 

Speaking of that, SolarFlare supports kernel bypassing technologies. If we can get Euphony or AudioLinux to take advantage of the OpenOnload technology, that would be a big winner! 

 

And if we want to go a step further, there are Solarflare AOE cards with a built-in FPGA. If someone wants to write the code for the FPGA, the network traffic would go from the network to the FPGA, completely bypassing the entire OS. But that is a hell of a task! And supporting that code would be a very time consuming task as well. Probably no one would go there. 

[Nenon]

BTW, I had a chat with Željko from Euphony. He compiled OpenOnload and is ready to work with someone who has a SolarFlare NIC and wants to try it. 

[Nenon]

On 2/5/2021 at 8:48 AM, StreamFidelity said:

The XILINX Solarflare Flareon Ultra SFN8522 network adapter was the reason to omit the cover, as the card needs a case fan with at least 200 revolutions for cooling. Without the cover this has been going on for weeks without any problems. But it was certainly not a permanent solution.

 

And so I thought about an acrylic glass lid, where the inside can also be seen. Voilà, that's it! And this is what the prototype looks like. This variant is temperature stable up to 80 °C. Note the practical cup holders. 😄

 

The circular openings are of course the ventilation holes. These are exactly above the CPU cooling (right) and the cooling for the PCIe cards (left).

 

The exciting question is, is that enough? Only four holes? But I already had a positive indication: the CPU temperature difference with the acrylic glass cover compared to the completely open chassis is only around + 3 °C (54-55 °C)! Now in 24h operation no malfunctions occurred and the low temperature was maintained. 👍

 

The prototype is 10mm thick and therefore looks very valuable. I didn't want a screw connection with this thickness. The lid is therefore loose.

 

I will design another prototype with screw connection and thin acrylic glass. Other variants are also conceivable. For example colored tinted acrylic glasses.

 

Very cool! 

[Nenon]

On 2/20/2021 at 7:11 PM, Darryl R said:

 

This is very nice work.  Larry said he was going to make a backplate without top and bottom edges this summer, but I thought he needed to lose those slots where the bracket bottoms snap in as well.  I'm so tired of taking that backplate off to mess with expansion cards.  I always wondered why he didn't do it this way to begin with.

 

If you redid the bottom plate, this might give the Taiko DIY chassis (which we haven't seen yet) a run for the money.

 

There is a very easy solution. Just cut the part in red:

 

[Nenon]

For those aiming at the lowest possible latency, why not try the Cisco Nexus 3548 (or 3524) switch. It's the industry standard for low latency (190-250 nanoseconds) switching without going to layer 1 switches. 

[Nenon]

What feeds the Taiko ATX is crucial. It makes a huge difference. Definitely worth trying the higher voltage unregulated LPS. 

[Nenon]

3 hours ago, StreamFidelity said:

Audio PC: passive cooling of Solarflare network cards

 

Due to the high performance of the Solarflare NICs (Network Interface Card) of the 8 and x series, the chips installed there get very hot. They are already equipped with an oversized cooler compared to other cards, but the manufacturer expects a case fan with at least 200 rpm. The only solution so far has been generous ventilation holes on the lid or cover up the lid.

 

The Solarflare NICs have very low latencies and are therefore used, for example, in high-frequency trading on stock exchanges. In the audio sector, low latencies are also a great advantage. Taiko Audio and others noted outstanding sound, but it deteriorated as the heating increased. That is also my observation. That's why I developed passive cooling with heat pipes.

 

 

The HDPLEX 2nd Gen H5 Fanless Chassis with heat pipes offers excellent passive cooling options by simply dissipating the heat via the side wall. If you don't use a passively cooled graphics card, the left side wall is free. Due to the different slots on the motherboards, a flexible connection option for the heat pipes was required. The attachment in the middle enables the heat pipes to be shortened. Since it is a prototype, the production is still a bit rough.

 

 

The block at the top enables it to be attached to the Solarflare NIC with four precisely fitting holes.

 

 

Here you can see the heat sink of the XILINX Solarflare X2 Series Ethernet Adapter - XtremeScale X2522, which is not sufficient for passive cooling. The contact pressure of the heat sink on the chip is created by means of spring plug-in holders.

 

 

The spring clips must be carefully removed from the back of the circuit board. The bracket is pressed together with tweezers and the connection to the circuit board is loosened.

 

 

The bare solar flare chip comes out and it is cleaned with isopropyl alcohol. Then a new thermal paste or thermal pad comes on.

 

 

The new cooling block is carefully attached to the chip with the removed spring plug-in holders.

 

 

It must be ensured that the cooling block lies flat on the chip.

 

 

A thermal paste must be applied so that the heat is dissipated well via the side wall. Before doing this, you will try out how the heat pipes are positioned and marked.

 

 

Installed it looks like this:

 

 

With the acrylic glass lid closed, the prototype is already impressive.

 

 

performance

 

I was very curious to see if that would work. After about an hour the left side wall gets quite warm, which is a very good sign. The heat dissipation via the outer wall works! An endurance test now follows. I will also completely close the housing cover with the original.

 

Sound impression

 

Even after hours there is finally no audible drop in performance. The enormous clarity, the stage far out the loudspeakers and the unexcited, slag-free sound are retained.

 

Well done! In a true DIY spirit. I like it! Thank you for sharing.