[bogi]

I'm using gauss-xla and there's a 5-6 second delay across the board. At 768k there are constant dropouts (3-4 per minute), at 384 almost none, but the lag is still there.

I'm hoping/guessing it's a RAM issue because my mini only has 8gb and I have 32gb on order. The mini is a 2018 with a 3.6Gb quad core i3, I would think this is plenty of cpu power for PCM?

I don't think more RAM is needed for xla. Rather faster CPU clock for 768k.
You may yet try other USB cable or port, or for example Topping HS02 galvanic isolator.
Try different 'Buffer time' in Settings -> Outputs and 'Blocks per cycle' in Settings -> Advanced.

 

[BShaw]

I don't think more RAM is needed for xla. Rather faster CPU clock for 768k.
You may yet try other USB cable or port, or for example Topping HS02 galvanic isolator.
Try different 'Buffer time' in Settings -> Outputs and 'Blocks per cycle' in Settings -> Advanced.

Thanks bogi. As I said, there is no USB connection in my chain, I'm running fiber to a converter then straight into my Gustard R26 with a high quality OCC silver LAN cable. Fiber is as good as it gets for isolation. I guess I could insert a LAN silencer but I don't see the point. I have one on the dirty end of things, maybe I'll throw it in and see what happens but I really don't think it will change the situation.
I'm also not sure 3.6g quad core is such a dog that it can't handle the signal, I have a buddy with a 2012 dual core 2.8 or maybe it's 3.0, and he has no issues with dropouts. The difference between his Mac and mine is that he has twice the RAM, so I figured this looks like the culprit.
Buffer time changes nothing, but I'll try the "blocks per cycle" setting later today, thanks for the tip. I work with my hands and was born before seatbelts and the only computers we saw were in films of NASA on the news. Some of this takes getting used to for a guy like me.

 

[Currawong]

Are you trying to add context or are you trying to educate Jussi on digital filtering?

Show me ringing on an over-sampled signal using a long sinc function with in-band content.

 

[Currawong]

I'm using gauss-xla and there's a 5-6 second delay across the board. At 768k there are constant dropouts (3-4 per minute), at 384 almost none, but the lag is still there. I'm hoping/guessing it's a RAM issue because my mini only has 8gb and I have 32gb on order. The mini is a 2018 with a 3.6Gb quad core i3, I would think this is plenty of cpu power for PCM? There are no USB connections in my chain, fiber from the network switch, and a very high quality silver LAN cable directly into my DAC from the fiber converter, also very high quality with LPS. Maybe this provides the missing info you referred to above in post 938 Bogi?

PCM filters don't require that much computing power overall. Is adaptive rate switched on? That being said, more than likely it's your Gustard, not the Mini.

 

[BShaw]

PCM filters don't require that much computing power overall. Is adaptive rate switched on? That being said, more than likely it's your Gustard, not the Mini.

adaptive rate is greyed, would it be better to check it? And yeah, I'm afraid you might be right, considering an Audio gd R8 and a couple of others for the near future. I'm finding it pretty hard to find a standalone DAC that competes the the Gustard below $2500US.

 

[jlaako]

I've been slowly piecing together why exactly that doesn't appear to be true in reality and dream of proving it mathematically one day

Since time and frequency have 1/x relationship, better you make one aspect, worse the other one gets. This can be also called (Fourier) uncertainty principle. Hearing is known to be able to beat Fourier transform in this respect. How good you can make both simultaneously is what I've spent enormous amount of time on, trying to get as close to the impossible as possible. This is one of the fundamental points in the DSP work I do.

If you only look things only through FFT spectrum analysis (frequency domain aspect), you may trend towards longest ones. But while doing so, you may end up with something that is really bad looked from time domain aspect. One needs to consider all dimensions of the same. This is actually a four dimensional problem.

From this perspective, both MQA (super short filters) and Chord (super long filters) are both right and wrong at the same time. This because they look at things only from one angle...

Essentially your mathematical proof is in that uncertainty principle.

 

[bogi]

A digital filter of sufficient length will not "ring" when fed with a bandwidth-limited (in-band) signal.

Digitally encoded sampled signal is not band limited. The signal represented by digital samples consists of infinite number of spectral replicas (images) of base band.

Picture source:

Do D/A conversion without proper filtering and you will get these images at analog output. They are best visible at NOS DAC output, when only slow DAC analog filter is in action, so that the filter is able to cut the nearest images only partially. Look at the 1st graph here, it shows what brings 0 - 22.05k sweep at DAC output in NOS mode (output scale is up to 5 MHz):

You see series of audio band images, their amplitude is lowered by shape of DAC analog filter. That's in reality what you called band limited signal! One needs to distinguish encoding of digital signal samples and the signal itself.

 

[bogi]

A digital filter of sufficient length will not "ring" when fed with a bandwidth-limited (in-band) signal.

A half year ago manisandher discussed your point of view (band limited signal vs. ringing) with Jussi on AS forum. Find more detailed discussion incl. some demonstrations on this page.

 

[jlaako]

PCM filters don't require that much computing power overall. Is adaptive rate switched on? That being said, more than likely it's your Gustard, not the Mini.

In such cases, if network through NAA is involved, first thing to check is whether 802.3x Flow Control is active. This is something that is negotiated by hardware at low level. Most NIC drivers have this enabled by default, but it also need to be accepted by the switch/router. Unmanaged switches practically all support this and have it enabled. While smart (configurable) switches have it either disabled or enabled by default, with some 50/50 ratio or something. So if one is using a smart switch, it is important to check it's settings.

Also common source of 802.3x not being active are some optical SFP modules that don't negotiate this capability.

Typical NAA has a low power ARM CPU, which is not able to handle sustained 1 Gbps speeds. Such cases can lead to network traffic stalls that cause drop-outs if 802.3x doesn't work. This is handled at hardware level, when the recipient sees it's hardware packet buffer gets full, it asks sender to pause sending data using the 802.3x flow control pause frames. When there's more space, the hardware will again tell the sender that it can accept more data. If this is not active, recipient's hardware buffer can overflow resulting in packet loss and re-sends which in turn increases traffic which makes situation worse. Faster the network, more important this is. For example NAS with 10 Gbps interface may not be able to write data that fast on a HDD.

NAA v5 has some protocol changes to improve situation in cases where such flow control is not functional. But this is essentially a hardware level issue. For example on the CPU used in Rendu and few other devices, the on-chip local bus link between the 1 Gbps NIC and CPU is slower than 1 Gbps (about 400 Mbps), so the situation needs to be handled at hardware level, before the CPU. 802.3x solves this at NIC level.

 

[jlaako]

On my Holo May HQPlayer cannot set the dac to DSD mode if it is enabled as an output device in Roon. Disable the ASIO driver in Roon and you can play DSD.

This is "feature" of Thesycon's USB drivers and ASIO driver, which goes to some kind of non-exclusive mode when multiple applications access the ASIO driver. And this drops out the DSD support.

But this is documented in the official Roon documentation, last bullet point here:
https://help.roonlabs.com/portal/en/kb/articles/hqplayer

ASIO is supposed to be exclusive access driver, only one application should be able to access it at once.

 

[jlaako]

HQPlayer can only do volume leveling for local music libraries, but I'd be happy to be proven wrong

It can do it also when you play natively from Qobuz. Unfortunately Roon doesn't use HQPlayer's volume for that (it could). And UPnP is just incapable of such things.

 

[jlaako]

Indeed for volume leveling from Qobuz you need to use Roon.

No! It is natively supported by HQPlayer!

 

[jlaako]

Nearly every other filter -- apodizing or not -- does not share this trait. With those other filters, when you input a sample value x at audio stream position t, and you inspect that sample value t at the output (so awaiting the latency of the filter), the value is almost certainly not x anymore. What's better is another matter, but this in=out thing and only interpolating in-between is something particular to that closed form filter.

Halfband filters don't touch the original samples (which could contain correctable errors!). Those are passed through as-is.

That also means that any such filter is more or less leaky by definition.

 

[genefruit]

@jlaako
If one was to use PGGB to upsample files to 16fs/24 bit, in your opinion, would be the best SDM Nx filter(s) to play those files?

 

[jlaako]

There's a great misunderstanding about ringing in digital filters. What people see as "ringing" when an impulse response is run through a filter is the result of the effect of a non-audio signal -- basically a single, non-zero sample amongst silence. A digital filter of sufficient length will not "ring" when fed with a bandwidth-limited (in-band) signal.

Except that your typical source is not sufficiently band-limited. Plus it typically contains digital clipping, which looks like a non-bandlimited square wave. And if it is, it will be already ringing - as result of band limiting.

Don't mix idealistic theory with real life...

And the ringing and other errors already exists in your digital PCM recordings due to the (imperfect) ADC or mastering digital decimation filters. We can largely fix those errors with suitable apodizing filters.

I have demonstrated this already several times over past decades through full microphone - ADC recording chains.

HQPlayer's Apod-counter gives you a hint.