Jump to content

DIY Project High Performance Audio PC with high quality wiring

Recommended Posts

  • 3 months later...
1 hour ago, evalon said:

I believe reclocking is very effective, however, IMHO it needs to be done close to the DAC's (~IC's) clock input so as to avoid the influence of other logic circuitry in the transmission chain. In many cases this is a tricky thing to carry out in practice and would mean actively modifying the DAC.


It's already been done about a month ago, right now it's available from this seller in Japan






The most critical part is handled by McDual XO and McFIFO






They compared their Daphile dual PC setup (costing almost 1 million yen) to another average Intel Celeron-based laptop running Windows 10 with zero optimization, for whatever reasons they couldn't hear any differences between the two. Then they also swapped 3 different USB cables but once again everything seemed to sound the same


Oyaide Continental 5S V2

Asoyaji Audio (with aerospace wires from Gore)

Generic cable from 100-yen shops




Later on someone else tried that DAC again with the unpowered version of Sablon USB cable (2020 model) and finally there's only an extremely small improvement due to the absence of Vbus (+5V) connection






So that's more like comparing one of the best (unpowered) USB cable out there to several "unremarkable" powered USB cables, some of us might get super angry while others are very happy.

Link to comment
  • 4 months later...
19 hours ago, jabbr said:

The latest and greatest is NVMEof https://community.mellanox.com/s/article/what-is-nvme-over-fabrics-x , that's where Xilinx/Solarflare and NVidia/Mellanox are duking it out. 


For Infiniband, are stuff like zero-copy and OS-bypass etc. only good for reading from / writing to (RDMA send / RDMA receive) storage devices?


In other words, we'll only achieve near-zero CPU utilization of Native-IB whenever we're transferring files between two Mellanox NICs





For something else like Roon Core + Bridge or HQP + NAA etc. could only go through IPoIB as follows, therefore it might not be all that meaningful in terms of CPU utilization if I weren't mistaken




Since IPoIB provides a normal IP NIC interface, one can run TCP (or UDP) sockets on top of it. TCP throughput well over 10 Gb/sec is possible using recent systems, but this will burn a fair amount of CPU. To your question, there is not really a difference between IPoIB and TCP with InfiniBand -- they both refer to using the standard IP stack on top of IB hardware.


The real difference is between using IPoIB with a normal sockets application versus using native InfiniBand with an application that has been coded directly to the native IB verbs interface. The native application will almost certainly get much higher throughput and lower latency, while spending less CPU on networking.



Another question about NVMe over Fabrics versus NVDIMM (e.g. Intel Optane DCPMM) versus creating a RAM drive manually




But this is all housekeeping, and doesn’t take full advantage of the NVDIMM’s speed advantage over an SSD or HDD.  More software needed to be developed to prevent NVDIMM accesses from being bogged down by slow I/O routines that were developed for HDD and SSD.  At the time that they were written these routines were significantly faster than the drives themselves.  With NVDIMMs the opposite is true: The storage hardware (the NVDIMM) is significantly faster than the I/O routines accessing it.


Where do the I/O routines of RDMA (Native-IB) actually stand? Maybe much closer to slow I/O routines that were developed for NVMe SSDs?



64GB DDR-3 LRDIMM ain't worth THAT much now, that means we could very well build 1TB RAM-based diskless NAS with one of those HFT motherboards from Supermicro







'Hyper-Speed' HFT servers with up to 30% lower latency


Link to comment



Arista Networks (NYSE:ANET) today announced the Arista 7130L Series, the next generation ultra-low latency, high-precision network application platform, with deterministic 5 nanosecond switching and virtually undetectable jitter.


Measuring the latency of a 4ns switch



We have conclusively shown that the latency of the Arista 7130-16 is very close to 3.95ns per hop.



Hmm, are these numbers actually relevant or otherwise?




5 ppb stability OCXO with low noise power supply replacing the motherboard system clock




The use of a 3ppb OCXO clock running directly at 24MHz and connected via a board track just a couple of inches away from the USB chip.




Ships with OCXO (0.2ppb) by default (<15us holdover/24 hours):




Exablaze ExaNIC GM 2 Port Network Interface Card. MFG#: EXANIC-GM CDW#: 5117269. $2,746.11. Advertised Price. Call. Empty Star.


Exablaze ExaNIC GM 2 Port Network Interface Card. MFG#: EXANIC-GM CDW#: 5117269. $2,752.99. Advertised Price. 11-13 Days.









If only we could pay extra for a direct connection to TIDAL's CDN servers hosted by Amazon CloudFront etc.




RF networks began to appear a little less than a decade ago as network providers competed to have the fastest route between Chicago and New York.


The situation increased the incentive for RF development, and soon after the first microwave networks were deployed, offering low-latency traders much faster round trips than even the quickest available fibre optic routes.


Link to comment
  • 2 months later...



These cards are hardware identical, differring only in what features are activated in the factory. To work out the different model number, you would need to check feature activations (e.g. by using "sfkey") From the serial number, however, I can see this is a -Plus variant card. Specifically: SFN8522-Plus XtremeScale 10GbE server I/O adapter with LL firmware, Onload license and PTP license




1. To use Onload, you do need a card which has Onload AppFlex.  The "-Plus" variant of the adapter does include this, so no additional purchase is needed.

2. You can check "onload --version" to check Onload is installed and running; and you can use either "sfkey" (available in our utilities package) to check which adapters have Onload AppFlex, or "onload_mibdump hwport" to check which currently have Onload enabled.  (as some could have been manually disabled).




To use TcpDirect you need a card which supports low-latency operation; which the "Onload" versions of cards include.  So this card should support TcpDirect.

(However, the 'Scaleout Onload' version available on base adapters does NOT include this.)


Solarflare SFS-OL Onload AppFlex License For SFN7000 And SFN8000 Series Adapters = $719.96


That's the whole point of getting another factory-enabled card on eBay instead of paying for a license.

Link to comment

BTW, I guess that both Windows and Linux could be somehow affected by some kinda balancing act among audio latency / data buffer / power consumption / CPU utilization / noise etc.




1. Wouldn't it be better, if all applications use the new APIs for low latency? Doesn't low latency always guarantee a better user experience for the user?

Not necessarily. Low latency has its tradeoffs:

  • Low latency means higher power consumption. If the system uses 10ms buffers, it means that the CPU will wake up every 10ms, fill the data buffer and go to sleep. However, if the system uses 1ms buffers, it means that the CPU will wake up every 1ms. In the second scenario, this means that the CPU will wake up more often and the power consumption will increase. This will decrease battery life.
  • Most applications rely on audio effects to provide the best user experience. For example, media players want to provide high-fidelity audio. Communication applications want to minimum echo and noise. Adding these types of audio effects to a stream increases its latency. These applications are more interested in audio quality than in audio latency.

In summary, each application type has different needs regarding audio latency. If an application does not need low latency, then it should not use the new APIs for low latency.


Ideally we're "outsourcing" at least part of the heavy lifting stuff to something like DPU as I mentioned here



Link to comment
  • 4 months later...

FYI - here's an official SFP28 from Solarflare and its model SOLR-SFM25G-SR-LL could be found under the "Supported Transceiver" section









Therefore -LL should actually stand for Low Latency and just wondering if there were anything special about that particular model? Solarflare was taken over by Xilinx while the archived version (dated November 2019) of the original site didn't seem to have SOLR-SFM25G-SR-LL (UPC # = 816191010991) listed anywhere





And then there's a similar one from Axiom SOLR-SFM25G-SR-LL-AX (UPC # = 840177844270) while that costs about the same






$730 for a pair of SOLR-SFM25G-SR-LL, obviously they better be able to make a significant difference for audio purposes. Otherwise, why would anyone bother to pay THAT much while we could get these affordable cards from this seller with 100% positive feedback?






Those cards are indeed coming from China, though.

Link to comment
  • 4 weeks later...
  • 3 weeks later...



Since Thunderbolt is allowed almost straight access to the CPU, it is able to lower the “round trip” audio latency from about an absolute minimum of about 4.5 ms through USB, to under 1 ms.



On 5/21/2021 at 9:24 PM, littlefooch said:

In the interim, as the Win10 releases have matured, and in particular in the Insider releases, W10 does provide support for thunderbolt networking.  This can mean a lot of different things.  What I'm testing in particular is the ability to use existing Win10 capability to bridge two ethernet adapters (one or two of which can be Thunderbolt Networking adapters).


This is becoming increasing more stable as the W10 releases mature.




So turns out you can do just that but I'm not sure even Asus realises you can!

I had the same question, I'd heard some people had got TB3 cards to work by bridging a couple of pins on the header cable instead of plugging it into the motherboard. Thought I'd try it.

Purchased the Asus Thunderboltex 4, plugged it into a dark hero motherboard (x570 with no thunderbolt header). To my surprise it worked. No header cable required.

Just the USB2.0 cable plugged into the motherboard and card in the last PCI-E slot, nothing else. Not drivers needed (or at least Windows went and got them for me if there were).

I now have my GPU into the Thunderbolt card, that goes into a thunderbolt HUB (because you can have those with TB4) then a thunderbolt to DisplayPort cable that goes into a my Samsung Odyssey G9 doing 5120x1440@240hz. UFO Test confirms it all working smoothly with no stutter.

Only thing not working if I can't install the Intel thunderbolt software but I have no idea what it does or why you would need it?

I honestly can't work out why Asus doesn't shout this from the rooftops!


So two cards would cost $230 and the latency of Thunderbolt Networking over an optical cable could be quite interesting, *IF* that were actually confirmed by Corning to be OK





$360 for each cable





Link to comment
  • 3 weeks later...

I didn't realize that we've got (at least) two separate XOs on X2522 until I saw your pictures. Pretty easy to tell that X4 should be 20MHz but it's quite tricky to see those numbers of X2.


Since an internally-powered SFN8522 was able to beat the externally-powered JCAT NET Card FEMTO + OPTIMO 3 DUO combo with ease, replacing some clocks on X2522 should elevate its performance by another notch or two IMHO.


BTW, it's interesting that X2522 and X2541 were recommended specifically for their KumoScale software





Solarflare Communications XtremeScale SFC9250 10/25/40/50/100G Ethernet Controller OR Solarflare XtremeScale X2541-100G Adapter






KumoScale software manages storage devices at the media level to deliver the best possible latency at all times.




Testing conducted in a networked environment that included the NVMe-oF specification and a TCP/IP transport showed that KumoScale software’s read performance is 12x faster than Ceph, and read latency is reduced by 60%. With similar loads, under the same testing conditions, write performance is 60x faster than Ceph while latency is reduced by 98%.




Our native mapper function uses NVMe-oF to provision block volumes over the network which provides the lowest latency and best performance results.


StarWind NVMe-oF Initiator for Windows

Link to comment
  • 2 months later...
On 6/29/2021 at 4:36 PM, StreamFidelity said:

Tips for refining music files on the hard drive with resampling


FYI - Bug head Resampler 1.01 is included in the latest package




Bug head Resampler 1.01 正式リリース&終了バグ修正



Link to comment
  • 5 weeks later...



Over time I have observed that the audio quality of the PC is related to the number of processes that run on it and to the latency.




I have disabled all network services, and now 20 base processes run, 19 excluding taskmgr.exe.
The sound has improved a lot, it is more precise, with more relief, and it is more focused. The timbral is very real.
Very grateful to all the suggestions of this forum




Now I have a very clean and detailed sound, with a three-dimensional soundstage and very well focused vocals and instruments.
It is much better than with Server Core 2019 loaded in RAM.


If we weren't interested in music streaming services, than it should be hard to beat something like that.


Even though we're going completely offline for the best SQ available, we could still keep our (relatively expensive) network switches for improving USB outputs with either Startech USB2G4LEXT2 or Icron Ranger 2304GE-LAN as follows







And I'd like to give some credits to @Nenon here since his post on WBF really played an important role




What if I disable all network cards in Windows, so the OS cannot perform any network activity? And what if I use HQplayer with no oversampling and connect via KVM to control the software. Well, this my friends was one of the biggest improvements on my DIY server.


Even though he's joining that particular family mentioned above while his plan to publish a guide for tweaking Windows 10 Enterprise LTSC 2019 etc. might be put on hold indefinitely, I'm still very grateful that he's been sharing other findings with all of us



On 12/19/2020 at 1:15 AM, Nenon said:

You should try Windows LTSC and ProcessLasso. 

I will be publishing a guide for tweaking Windows for this motherboard sometime next year. Planning to share everything I've learned since April. I am also working with Emile from Taiko on testing some of his technology that will most likely treacle down to the DIY market sometime next year. This coming year will be very exciting for DIY. 



On 1/12/2021 at 2:49 AM, Nenon said:

I wish I could do that, but I can't find a way to do it legally. The current plan is to provide:

a. Instructions on how to obtain Windows LTSC. Someone else is working on that.

b. A configuration file for Ntlite to remove a bunch of components.

c. Some scripts that you can run to optimize Windows.

d. Some additional configurations.

e. A Process Lasso baseline config. 

That should do as a start. After that I expect people to tell me how to improve my server :). 


So here's our chance to create something truly special because IMHO either Win10 PE or Win11 PE (especially if we're preventing all network services to be started in the first place) could very have the potential to beat a stripped-down version of LTSC.



BTW, I posted something about Intel AMT / vPro below (VNC Server is ALWAYS on even if all network cards were disabled by the operating system) and it could be relevant for Alder Lake




Intel roadmap leak shows Alder Lake vPro arriving in 1Q22


Link to comment

Create an account or sign in to comment

You need to be a member in order to leave a comment

Create an account

Sign up for a new account in our community. It's easy!

Register a new account

Sign in

Already have an account? Sign in here.

Sign In Now

  • Create New...