Randy Myers
Well-known member
- Thread Author
- #1
We got off topic in another thread so I figured it was best to start a new thread as to not hijack the other one.
Norman wrote the follow:
The NAA will isolate the Roon server from the Dac.
SG Server uses Windows 10 and optimizes it for audio use. In Win10, the DLLs are "intelligent" and can be made to give priority to audio playback and upsampling computations.
Here is what you MAY miss compared to an expensive Sound Galleries Server (CompAudiophile forum):
The rollout has been delayed due to the need to redesign the case to handle the cooling loads which come from doing PCM to DSD 512 conversion. Along the way, we also upgraded the power supply to the motherboard which now has ripple in single digit microvolts
On the OS side we have made great strides in doing a Windows 10 build that has a small memory footprint and a low thread count
The OS has been set up so that HQ Player and Roon server have their own computing environments.
Another data point we have is Audiodan at the Italian Nexthardware forum. He has been running NAA for a longtime, and always told me that the spec and implementation of the NAA was very important. Audiodan has been running a DSD 256 self-built chipless DSC converter, on which he claims he can hear every little change upstream.
Another point to think about is we find NAA implementations sensitive to the power supplies used for the network switch, and the power supply used to power the fibre media converter when we have fibre optic isolation on the PC's Ethernet connection.
Our experience is that there is no free lunch when using Ethernet for audio, although for most installations it does work well.
=====================
Originally Posted by tranz
️ This is big news. Up until now NAA was touted as the way to separate the noisy computing environment from the delicate DAC side.
The NAA architecture is still valid and beneficial to most of us that use regular PC for the Hqp. If the noise could be handled efficiently in the PC (as does the SGM), it does avoid one hop processing and latency.
The SGM motherboard is still off the shelf board assembled with carefully attention to parts selection and various upgrades (clocks, power supply, bios tweaked, etc). The uRendu is a dedicated board and designed especially for audio. I think a bigger question to ask is how does the SGM sounds in comparison to PC + uRendu w/ high end power supply in Hqp NAA mode and doing DSD512.
=========================
10 months ago, we were getting very good results with NAA, but when we moved from a MOSFET power supply to a dual choke power supply for the motherboard, the performance of a NAA arrangement was no longer better than a one box solution, but actually inferior to a one box solution.
Our take is that with the lower noise floor which we achieve with the dual choke supply, the noise that comes with passing data over Ethernet becomes audible. Ethernet is a well isolated solution, but it does come with its data overhead, and the Ethernet receiver is not noise free either
===================
I would like to take the opportunity to give some of the background why the server costs what it does. Successful heavy computation for an audio application to sound good needs three things, control and suppression of RFI, dissipating about 50 watts heat from the CPU and maintains stable core temperatures, and improving the clock timing of the computation as much as possible.
- The mother board power supply is a big and heavy affair with 660,000 farrads of Mundorf Audio grade capacitors, over specified Cree Shottky rectifiers and 2 chokes which weigh over 3.5 Kgs each. The ripple from this arrangement is in the micro volts, not milli volts. When we started this project, we had a very nice mosfet regulated power supply that sounded great with a i7 4970. However as we upgraded to a Skylake, we were no longer able to get the mosfet supply to perform reliably with the power draw spikes of the i7 6700 K, and we had to start from scratch with a different PS concept. We decided to try out an old analog amp practice, a choke power supply design in a similar vein to Uptones JS-2 design, which we knew had great results powering Mac Mini's. We tried bigger chokes and more caps, and the sound got better and better. The last interation had a bill of materials increase of 300 euros, but the sound stage depth was way way deeper. What were we to do, keep the cost down or increase the price and deliver a soundstage with added realism ?
On the cooling side, the arrival of the T+A DAC 8 DSD doing DSD512, dramatically increased the computation and wattage of the CPU. The only fan less way to cool the CPU was to custom machine finned heat sinks from solid copper billet, very heavy and expensive, but cools very well. The cooler block we use on top of the CPU, is the same cooler block that is used in a German army battle tank, but we had to modify it so that we have individually adjustable pressure pads, again custom machined
on the SSD, it took us 4 months to find a power supply that sounded better than a Li-Po battery, and we commisisioned a supplier in Sweden to do a custom power supply with a very high frequency bandwith
to improve the timing, we removed the mobo clock and replaced it with first a TCXO, then 50 ppb, and then 5 ppb OCXO's. The 5 ppb OCXO easily sounded the best
and so it goes on
+++++++++++++++++++
At such high retail price, how does it compare to an Aurender W20 ?The W20 is one of the best, if not the best low power, small RF foot print ,bit perfect delivery server out there.
IMHO it's forte is delivery a very high quality SPDIF output, very RF clean, and a very stable clock. The Aurenda has a OCXO built in from the same supplier as one of the two that we have extensively tested.
The software player that comes with the Aurenda, and the only one that the W20 works with is based on the open source Linux MPD Player. Aurenda have done a commendable job of library management and developing tablet apps that work with their MPD player.
The digital music delivery of the SGM 2015 is completely different. The SGM has been designed, engineered, and built to employ computational intensive algorithms to convert all digital formats to DSD512. The reason why we have built this server is to provide the best computing platform for HQ Player to do its magic converting Red book lead into DSD512 Gold.
This weekend has been a highlight of our digital audio journey, we had the chance to spend quality time with Jussi Laako. There are some things I have learned that I want to share with forum friends and put on the record
Jussi Laako spent 7 years in the Finnish Navy. Part of Jussi's job was training sonar operators, an intensive 6 months after which the operators had any amazing ability to identify the sounds of different ships and submarines passing by. How many audiophiles have undergone 6 months of 5 days a week listening training LoL ?
Jussi's other job was to write new computational algorithms for processing sonar returns. Every couple of weeks Jussi would have a new version to test, HQ Player users, does this sound familiar ?
Then Jussi went on to work for Nokia. All cellphones use 1-bit sigma delta modulation. There are no PCM speach phones, have not been for decades !
Now Jussi works for Intel in Finland, I wonder why Intel has a R&D facility in Finland
Let me share with you, some computional bits with you. All of the processing stages have an output in 64 bit floating point format. To maintain the precision of the out put of some stages, Jussi actually inputs 80 bit floating point data, so that the output will have 64 bit precision. To do the 80 bit manipulations, Jussi has had to write his own assembler code. There are no compilers available in the public domain to handle 80 bit floating point.
HQ Player's processing of PCM and other formats is totally another world from the small footprint, bit perfect MPD world inhabited by Aurenda and Antipodes.
On the user interface front end, the SGM has as it's default software Roon 1.2. Roon absolutely is the busiest piece of media management software out there, which requires a heft amount of CPU power but comes with its RFI baggage as well. There is a reason why Aurenda have decided not to support Roon, and I can fully understand Aurenda's decision at this time.
By contrast, the SGM's sledge hammer approach to audiophile computing allows the running of both Roon and HQ Player at DSD512 in the same box. Not an easy task, just ask Lukaz of Lampizator who was taking pictures of the CNC machined copper heatsinks, on Friday night after dinner
enough for now, have to jump in the shower to make it for breakfast with the Roon guys !
My reply:
Thanks Norman. Appreciate the feedback. I can see how the high end Music Servers are more refined in their approach, which is cool. Some good incite there on optimizing. It has given me some ideas on ways of improving my server a little more even!
One other thing I realized is i was making another huge mistake. Since this is a headless server and no monitor is likely to be hooked to it there is no need for a video card. I didn't think about this and had left the huge GTX-980 graphics card hooked up. I just ripped down the PC and removed this card (a small tight case so not as easy as a full size case to get this large card out). This card had to be using a lot of power, but more importantly was probably creating noise and certainly heat inside the case... I have not had the chance to listen yet, but I did confirm that the machine boots and serves Roon files just fine after the video card extraction....
Norman wrote the follow:
The NAA will isolate the Roon server from the Dac.
SG Server uses Windows 10 and optimizes it for audio use. In Win10, the DLLs are "intelligent" and can be made to give priority to audio playback and upsampling computations.
Here is what you MAY miss compared to an expensive Sound Galleries Server (CompAudiophile forum):
The rollout has been delayed due to the need to redesign the case to handle the cooling loads which come from doing PCM to DSD 512 conversion. Along the way, we also upgraded the power supply to the motherboard which now has ripple in single digit microvolts
On the OS side we have made great strides in doing a Windows 10 build that has a small memory footprint and a low thread count
The OS has been set up so that HQ Player and Roon server have their own computing environments.
Another data point we have is Audiodan at the Italian Nexthardware forum. He has been running NAA for a longtime, and always told me that the spec and implementation of the NAA was very important. Audiodan has been running a DSD 256 self-built chipless DSC converter, on which he claims he can hear every little change upstream.
Another point to think about is we find NAA implementations sensitive to the power supplies used for the network switch, and the power supply used to power the fibre media converter when we have fibre optic isolation on the PC's Ethernet connection.
Our experience is that there is no free lunch when using Ethernet for audio, although for most installations it does work well.
=====================
Originally Posted by tranz
The NAA architecture is still valid and beneficial to most of us that use regular PC for the Hqp. If the noise could be handled efficiently in the PC (as does the SGM), it does avoid one hop processing and latency.
The SGM motherboard is still off the shelf board assembled with carefully attention to parts selection and various upgrades (clocks, power supply, bios tweaked, etc). The uRendu is a dedicated board and designed especially for audio. I think a bigger question to ask is how does the SGM sounds in comparison to PC + uRendu w/ high end power supply in Hqp NAA mode and doing DSD512.
=========================
10 months ago, we were getting very good results with NAA, but when we moved from a MOSFET power supply to a dual choke power supply for the motherboard, the performance of a NAA arrangement was no longer better than a one box solution, but actually inferior to a one box solution.
Our take is that with the lower noise floor which we achieve with the dual choke supply, the noise that comes with passing data over Ethernet becomes audible. Ethernet is a well isolated solution, but it does come with its data overhead, and the Ethernet receiver is not noise free either
===================
I would like to take the opportunity to give some of the background why the server costs what it does. Successful heavy computation for an audio application to sound good needs three things, control and suppression of RFI, dissipating about 50 watts heat from the CPU and maintains stable core temperatures, and improving the clock timing of the computation as much as possible.
- The mother board power supply is a big and heavy affair with 660,000 farrads of Mundorf Audio grade capacitors, over specified Cree Shottky rectifiers and 2 chokes which weigh over 3.5 Kgs each. The ripple from this arrangement is in the micro volts, not milli volts. When we started this project, we had a very nice mosfet regulated power supply that sounded great with a i7 4970. However as we upgraded to a Skylake, we were no longer able to get the mosfet supply to perform reliably with the power draw spikes of the i7 6700 K, and we had to start from scratch with a different PS concept. We decided to try out an old analog amp practice, a choke power supply design in a similar vein to Uptones JS-2 design, which we knew had great results powering Mac Mini's. We tried bigger chokes and more caps, and the sound got better and better. The last interation had a bill of materials increase of 300 euros, but the sound stage depth was way way deeper. What were we to do, keep the cost down or increase the price and deliver a soundstage with added realism ?
On the cooling side, the arrival of the T+A DAC 8 DSD doing DSD512, dramatically increased the computation and wattage of the CPU. The only fan less way to cool the CPU was to custom machine finned heat sinks from solid copper billet, very heavy and expensive, but cools very well. The cooler block we use on top of the CPU, is the same cooler block that is used in a German army battle tank, but we had to modify it so that we have individually adjustable pressure pads, again custom machined
on the SSD, it took us 4 months to find a power supply that sounded better than a Li-Po battery, and we commisisioned a supplier in Sweden to do a custom power supply with a very high frequency bandwith
to improve the timing, we removed the mobo clock and replaced it with first a TCXO, then 50 ppb, and then 5 ppb OCXO's. The 5 ppb OCXO easily sounded the best
and so it goes on
+++++++++++++++++++
At such high retail price, how does it compare to an Aurender W20 ?The W20 is one of the best, if not the best low power, small RF foot print ,bit perfect delivery server out there.
IMHO it's forte is delivery a very high quality SPDIF output, very RF clean, and a very stable clock. The Aurenda has a OCXO built in from the same supplier as one of the two that we have extensively tested.
The software player that comes with the Aurenda, and the only one that the W20 works with is based on the open source Linux MPD Player. Aurenda have done a commendable job of library management and developing tablet apps that work with their MPD player.
The digital music delivery of the SGM 2015 is completely different. The SGM has been designed, engineered, and built to employ computational intensive algorithms to convert all digital formats to DSD512. The reason why we have built this server is to provide the best computing platform for HQ Player to do its magic converting Red book lead into DSD512 Gold.
This weekend has been a highlight of our digital audio journey, we had the chance to spend quality time with Jussi Laako. There are some things I have learned that I want to share with forum friends and put on the record
Jussi Laako spent 7 years in the Finnish Navy. Part of Jussi's job was training sonar operators, an intensive 6 months after which the operators had any amazing ability to identify the sounds of different ships and submarines passing by. How many audiophiles have undergone 6 months of 5 days a week listening training LoL ?
Jussi's other job was to write new computational algorithms for processing sonar returns. Every couple of weeks Jussi would have a new version to test, HQ Player users, does this sound familiar ?
Then Jussi went on to work for Nokia. All cellphones use 1-bit sigma delta modulation. There are no PCM speach phones, have not been for decades !
Now Jussi works for Intel in Finland, I wonder why Intel has a R&D facility in Finland
Let me share with you, some computional bits with you. All of the processing stages have an output in 64 bit floating point format. To maintain the precision of the out put of some stages, Jussi actually inputs 80 bit floating point data, so that the output will have 64 bit precision. To do the 80 bit manipulations, Jussi has had to write his own assembler code. There are no compilers available in the public domain to handle 80 bit floating point.
HQ Player's processing of PCM and other formats is totally another world from the small footprint, bit perfect MPD world inhabited by Aurenda and Antipodes.
On the user interface front end, the SGM has as it's default software Roon 1.2. Roon absolutely is the busiest piece of media management software out there, which requires a heft amount of CPU power but comes with its RFI baggage as well. There is a reason why Aurenda have decided not to support Roon, and I can fully understand Aurenda's decision at this time.
By contrast, the SGM's sledge hammer approach to audiophile computing allows the running of both Roon and HQ Player at DSD512 in the same box. Not an easy task, just ask Lukaz of Lampizator who was taking pictures of the CNC machined copper heatsinks, on Friday night after dinner
enough for now, have to jump in the shower to make it for breakfast with the Roon guys !
My reply:
Thanks Norman. Appreciate the feedback. I can see how the high end Music Servers are more refined in their approach, which is cool. Some good incite there on optimizing. It has given me some ideas on ways of improving my server a little more even!
One other thing I realized is i was making another huge mistake. Since this is a headless server and no monitor is likely to be hooked to it there is no need for a video card. I didn't think about this and had left the huge GTX-980 graphics card hooked up. I just ripped down the PC and removed this card (a small tight case so not as easy as a full size case to get this large card out
). This card had to be using a lot of power, but more importantly was probably creating noise and certainly heat inside the case... I have not had the chance to listen yet, but I did confirm that the machine boots and serves Roon files just fine after the video card extraction....
