DSD performance issues

Here it is

Looks like the convolution has to process a lot less taps in this mode.

1 Like

Thanks. That’s what I thought. The filter is applied at 352.8 kHz, not 5.6 MHz. That’s probably why it only requires 373k taps vs. 5968k. The ratio of tap count is basically the same as the ratio of frequencies:

5968k taps / 373k taps =16
5,644.8 kHz / 352.8 kHz = 16

That’s where the perf difference is coming from. Saying that processing at DSD rates is overkill is an understatement.

1 Like

Cool, problem solved. Avoid downloaded DSD files if you do not have a blindingly fast cpu to process it.
Convolution filters are not your friend in this scenario.

1 Like

If you really really want DSD, you can try to set up Roon to down-sample to 352.8, apply DSP, then up-sample back to DSD. (Not sure how, I guess by disabling DSD processing?) However, in my book, DSD only makes sense if sent directly to the DAC, without processing, since its only purpose is to be “D/A-friendly”.

As several others have pointed out, the problem is the massive number of taps. I use HAF convolution filters as well and have the same problem. My system will start out a native DSD file at .6X and over the course of a second or so spins up to 1.4X and the file starts playing. Like you, there is no problem with files upsampled to DSD, just with native files. The solution I use is to have Roon convert the files to their equivalent PCM rates. I know you said you don’t want to do that but I’m not sure where you’d get a computer with enough grunt to do what you’re asking of it.

Apparently, Linux is the OS of choice these days - should be a doddle to get Roon Core to run on one of these :grin:

1 Like

For the price of a Nucleus+, you can get a PC that is more than capable of doing that. It would still be overkill though.

Those supercomputers are all distributed. You need single system performance :slight_smile:

According to the article (and my understanding of the term), they are all non-distributed computers.

As far as I know all current high end supercomputers are distributed.
That is, N computers, each running the same OS, using software to distribute compute jobs across them.
None of the current ones are just a single computer with lots of CPUs in it (as that doesn’t scale very far).
I’ve visited this one btw:

Note it says " 4,608 compute servers, each containing two 22-core IBM Power9 processors and six NVIDIA Tesla V100 graphics processing unit accelerators, interconnected with dual-rail Mellanox EDR 100Gb/s InfiniBand."

1 Like

Seriously cool. Makes my Roon Core setup look tiny by comparison:

That’s pretty impressive as well!
Now we just need roon to make a distributed version so we can run it on a supercomputer and be able to scan real fast. Not to mention play dsd 512s :slight_smile:

1 Like

I bet they’ll start selling DSD4096 soon after.

Roon DSP for a single zone uses a single thread, but the particular step of Sigma Delta Modulator can two use threads (if Parallelize Sigma Delta Modulator) is enabled.

That is why you need a high single thread performance for Roon DSP.

1 Like

Yes. This is precisely why Native DSD Processing needs to be disabled for this use case.

DSD256 with convolution does just do it at 1.1x here without stumbling.

But I probably don’t have an adequately resolving system :wink: to hear the difference…
Oh wait, I’ve got music scope to look at the contents of the file…

Okay, I think it’s just the hype about who’s got the larger numbers…

1 Like

To add to that:

That looks like a very cool tool (musicscope)! Tried to download it from sourceforge but fails to run on the mac. Sent them an email and hopefully I can get a working version…

Got it running. Very slick!

I got the tool, but it looks like it doesn’t show any results when analyzing 32bit int or float samples, although it behaves as it’s going through the file. It seems to be capped at 24bit… Weird.
P.S. Skimmed through the manual and that’s expected. Still weird.