DIY Project High Performance Audio PC with high quality wiring

[MarcelNL]

des Rätsels lösung is close...I suspect the SFP I'm using is a multi mode...cisco glc-sx-mm 30-1301-02...how it initially seemed to work is a mystery to me, but now I know I need to wait for the Finisar's to arrive🤐

[drjimwillie]

Good day Audio Friends,
I hope you are enjoying some fine music🎶
 

Has anyone tried the new Jcat net card XE? I know that they shipped, but I don’t see anyone talking about them.

 

I would really love someone to do a head-to-head comparison of the Jcat net card XE to the solarflare card.

[MarcelNL]

an epiphany, the batch of Finisar SFP 1321 BTL I last ordered arrived first, after a lot of juggling around it now works. They are currently warming up and starting their break in...tonight when I'm done roasting some coffee I'll do a quick listen. It's hard to compare against a week or longer ago with the multimode SFPs but I trust that a large difference will be noticeable. 

[MarcelNL]

Of course I was unable to wait, first impression is good!

Far better than with Cat 6, but that was already clear earlier on.

 

A first impression is that the background is inkier, more dynamics, slam, pace etc are all in a good place, low end is tight, but overall resolution is the big winner !

 

[MarcelNL]

after a longer listening the first impression is confirmed but in addition to that it seems there is an overall sort of decompression effect going on. Where the low end in my system left little to the imagination definition in the (ultra) low end got better.

I LIKE what I'm hearing a LOT, especially knowing this is streaming from the internet (Tidal, and I know Qobuz is better...I will switch to Qobuz once the Tidal subscrition runs out in a few weeks) !

 

 

[Exocer]

Xilllinx vs Startech to commence tomorrow. Will let some tunes play overnight for initial break-in. So far, it is sounding at least as good as the Startech.

 

Will start with my favorite SFP: Finisar 1321

 

[MarcelNL]

@exocer

Is the result so close that you're still A-Bing or are you so baffled with the sound ? ;-)

(just curious about what your findings are)

 

 

[Exocer]

6 hours ago, MarcelNL said:

@exocer

Is the result so close that you're still A-Bing or are you so baffled with the sound ? ;-)

(just curious about what your findings are)

 

 

@MarcelNL 

They are very different from each other. Both provide low noise floor but the Xilinx is a bit better in this regard. The Startech sounds a bit more musical and the Xilinx is more clean/precise so far. Lets give the card more time to break in. I did switch to my Planet SFPs to clean my pallet and because I did not like the 1321s with the Xilinx out of box. I will switch the Finisars in today as the sound of card is changing for the better as time goes on.

 

Note: I use a larger and brand new adapter for the Xilinx card so that too may require some settling in. 

 

Will give further updates as things progress.

 

[MarcelNL]

Sweet, thanks for satisfying my curiosity!

 

The Xilinx needs a full size PCI slot, my MB only has 2 of those...I may have to stick the PF card in the 1x slot when the Xilinx gets here...

 

 

[StreamFidelity]

50 minutes ago, Exocer said:

Note: I use a larger and brand new adapter for the Xilinx card

 

I think a riser cable is very unfavorable. I prefer a direct connection.

 

All eight PCIe lanes are used for data connection to the CPU, so a slot with a direct connection to the CPU is recommended.

 

RSS uses all 16 logical processors of my i9-9900K CPU. Activation of hyper-threading is required!

 

There are other recommendations:

 

- Interrupt moderation
Disable in Network Adapter’s Advanced Properties.

 

- Max frame size
Configure to maximum supported by network in Network Adapter's Advanced Properties.

 

- Receive Side Scaling (RSS)
RSS is enabled by default for best networking performance. It is best to set the maximum number of RSS queues to be equal to the maximum number of RSS CPUs.

[Exocer]

17 minutes ago, StreamFidelity said:

 

I think a riser cable is very unfavorable. I prefer a direct connection.

 

All eight PCIe lanes are used for data connection to the CPU, so a slot with a direct connection to the CPU is recommended.

 

RSS uses all 16 logical processors of my i9-9900K CPU. Activation of hyper-threading is required!

 

There are other recommendations:

 

- Interrupt moderation
Disable in Network Adapter’s Advanced Properties.

 

- Max frame size
Configure to maximum supported by network in Network Adapter's Advanced Properties.

 

- Receive Side Scaling (RSS)
RSS is enabled by default for best networking performance. It is best to set the maximum number of RSS queues to be equal to the maximum number of RSS CPUs.

 

Thanks! I keep my i9 9900k Hyper Threading activated at all times. My server runs on Euphony so I am not sure how to apply some of the recommendations but will look into it.

 

Yes, the more I think about it the more I realize this build deserves a full ATX board. That will be in the cards soon.

 

For now, it is using a riser cable. Perhaps the Startech is a better chose for those with Mini-ITX boards...maybe not. Time will tell. 

[Exocer]

I've switched the Finisar 1321's back in (from the Planet tech MGB-TLX) and today the card sounds A LOT better. The very first observation I didn't share is that the midrange seemed a bit recessed with the Xilinx out of the box and now they're coming in beautifully, like the Startech. I am really liking the sound today.

 

The plan is to leave the Xilinx/1321 combo in for a few weeks and switch back to the Startech.

 

I believe I have reached the point where the strengths of the Xilinx are apparent while at the same time giving me what I liked about the Startech. So this could be considered a win (so far).

[MarcelNL]

@StreamFidelityinteresting !

 

Is best networking performance needed for day to day audio streaming? I mean, can the difference be heard ? Just wondering, 1Gbit or even 10Gbit (depending on which FSP module is used) is much faster than a what a 'regular' stream would need right?I'm not referring  to server to rendered traffic using f.e. DSD512 data streams, but more regular Tidal or Qobuz streams. 

16 threads required for a bit of networking almost sounds as if there is a lot of overhead.

 

[MarkusBarkus]

...I recall some folks were speculating that because the criteria/spec for 10Gbit was so much tighter, that it suggested the supporting hardware was "better" and perhaps quieter. I don't recall that we ever conclusively established that as factual. 

[StreamFidelity]

3 hours ago, MarcelNL said:

Is best networking performance needed for day to day audio streaming? I mean, can the difference be heard ?

 

Define best networking performance. What's the best? Jitter-free? Free from noise? High bandwidth? Speed? I think we want a bit of everything. Reclocking also reduces jitter and the lowest possible latencies. 

 

That's why I use media converters and switches modified with special Sellarz Clocks. Others use EtherRegen or something similar. Solarflare 8522 has a Stable Precision Oscillator Stratum 3 compliant.

 

3 hours ago, MarcelNL said:

1Gbit or even 10Gbit (depending on which FSP module is used) is much faster than a what a 'regular' stream would need right?

 

Glass fiber is the perfect galvanic isolation and blocks the high frequency noise. 10GBit/s are better than 1GBit/s are better than 100MBit/s. Some say because there are more reserves and higher data bursts are possible. For DSD and NAA I need 1GBit/s, because with 45GBit/s half of a 100MBit/s connection would be occupied. I talked about the bandwidth and not about the speed!

 

3 hours ago, MarcelNL said:

16 threads required for a bit of networking almost sounds as if there is a lot of overhead.

 

Speed is expressed in latencies. In other words, the ability to process input and output in a very short time. And with that I come to the end of my short presentation. The Solarflare card is used in high-frequency trading on the stock exchange, where a decision is made between profit and loss in a split second. We also need the lowest possible latencies for our audio PC, because this drastically reduces jitter. Less jitter means in audio more SQ.

 

In order to keep the latency low, RSS over 16 lines is so important, because 16 x parallel data processing is possible. Therefore, there are sometimes problems with the temperatures, because the Solarflare FPGA chip has a lot of computing power.

[StreamFidelity]

12 hours ago, StreamFidelity said:

For DSD and NAA I need 1GBit/s, because with 45GBit/s half of a 100MBit/s connection would be occupied.

 

A typo, it should read 45MBit/s. 

[MarcelNL]

Thanks for that extensive response!

 

By Best networking I intended to write, best sounding networking, in the end that is the goal and intended use case. For my purposes DSD streaming is not an issue as I prefer PCM, I understand that the bandwidth required for DSD 512 is considerable yet still nowhere near 1Gbit unless I'm wrong;  - I got that 45 Gb/s was a typo ;-0 )

 

 

Using fiber for a few days now I can relate to the advantages it brings (it's similar to using I2S as digital 'interlink' vs USB), using better clocks is a clear winner in my experience too. I can not however attribute the advantages to any one of fiber's technical aspects, I suspect galvanic isolation may have something to do with it but with my experiences in EMI shielding CAT 8.1 cable it might also have to do with an absence of whatever is affecting copper wire (even STP).

 

I certainly see potential merit in the type of card such as the Xilinx I have underway, for one in deleting one more conversion step and a switch (conversion is bound to have effects and the switch is likely inherently less precise than the fibers and SFP's capabilities) and secondly given it's precision clock it should be lower jitter within the realm of the fiber connection. The card IMO will not affect latency within the computer as that is another domain driven by other processes and principles or I'm understanding it incorrectly. Would be interesting to use the card with the external high precision clock, as from what the spec sheet says the latency goes down to dazzlingly low numbers, and hear what that does.

 

Will loading the CPU with all this network activity have an effect on it's capability and affect SQ was the thing I wondered about mostly, driven by what folks do to tame W10 doing better audio by cutting down all unnecessary processes? Less is usually more in audio, less conversion, less solder joints, less elements, less cable. I'd like to be able to use just enough of everything for best SQ, but not more than that. Seeking an optimum is difficult, I know, it was more of a philosophical question as I suspect there are few who can answer to this. 

  

 

 

 

 

 

[Exocer]

Update on Euphony OpenOnload capability:

 

Zeljko had some issues setting up OpenOnload on my system remotely. Also, losing connectivity and an inability to see errors makes it difficult for him to work. We will need to wait for him to obtain the card in-house to test/deploy the proper items and push it out to the rest of us. Will be looking forward to that :)

[MarcelNL]

look what the mail brought me today, now to find out if it works from scratch in Daphile or not 

(after I figure out how to accomodate three quite thick PCIe cards in the three slots available on the MB 

 

 

[StreamFidelity]

@MarcelNL  Congratulation. Check out Linux drivers here:

 

https://support-nic.xilinx.com/wp/drivers

[MarcelNL]

thanks guys, I'm, afraid however that I need to follow the Linux for dummies part A training when Daphile does not support it straight away...I have a fundamental understanding of how it all works but have never worked with Linux/Unix...so sorting out adding a driver to a packaged OS may be beyond me for a while...

Will report on where I am after installing the card. 

edit: no dice....the three cards don't have enough space.

 

Given that the PF is approx two slots wide and the graphics card and Xilinx card need a full size PCIe slot the PF I2S will likely need a riser to work on MB and case...

So as all shops are closed it'll be a couple of days until I can test this after mail ordering a riser.

[MarcelNL]

will try sticking the graphics card (Daphile works headless anyhow but the MB needs the GPU to boot) in a 1x to 16x riser, so the PF and Xilinx cards can have a proper slot...we'll see..

[MarcelNL]

received the riser card today, only to discover it has at least two damaged circuit board/connector traces of the sort that likely never will come out of the connector on the MB once inserted/jammed in there. A new one is on its way.

An interesting feature of this riser card is two capacitors of 100uf on each side of the 16x adapter, likely to buffer power compensating for the lead length, something I wonder about adding to the power cables from the PSU to the MB. 

I have a hunch that adding some very nice NOS specialty caps may do wonders for silencing noise picked up underway and at the same time provide a small buffer for power peaks. My experience so far with EMI shielding of the power cables is good, there must be more to be gained in that segment.

[Exocer]

@MarcelNLdid you purchase the "onload" version?

[MarcelNL]

7 hours ago, Exocer said:

@MarcelNLdid you purchase the "onload" version?

@Exocer I don't think so, is there a way to distinguish between the versions? I got the one showing in the picture showing a bit earlier in this thread.

 

reading the leaflets it seems that the difference is software being enabled, likely by programming an eeprom different or something, so the main difference is probably only in numbering showing on the circuit board.