I’ve been listening to poly-sinc-xtr-mp-2s and AMSDM7 512 fs+ for a week or so now, upsampling to DSD 512 into a microRendu NAA and then to a Holo Audio Spring set to NOS. The Spring uses a discrete resistor ladder completely separate from PCM for DSD.
I have to say I’ve fallen in love with HQP all over again as a result of this filter and modulater combo. I had been listening to PCM for a while and was preferring it to closed form DSD, which had been my favourite. PCM had weight and punch that was missing.
But the xtr filter has brought that weight back and added top end sparkle without a trace of sibilance.
So can you tell us a bit more about the xtr filters ? In what ways are they similar to or different from closed form ? Why are they so computationally intensive ?
I haven’t been able to use the non-2s version of xtr-mp with a 7700 Kaby Lake and a GTX 970 when upsampling Redbook to DSD 512. I’ve tried both with and without multi-core and CUDA and still get stuttering. What spec machine would be needed to do that ?
And thanks all over again. This combo makes me sit and listen when I’m just walking past.
Heck I find the xtr filters great when upsampling to 352/384 PCM as well.
Oddly DSD128 not so much, DSD256 things shine again. But my DAC has the worlds worst drivers and I can’t use DSD256 without fear of blowing up my speakers. So I’m stuck in PCM for now.
Thanks Jeff, Yes, I have the limit set at 512 x 48. Haven’t tried setting auto-rate. I read on CA that people were having more success with non-2s by disabling CUDA, but I think that was a few upgrades ago.
I’m not planning on upgrading the Server at the moment, very happy with 2s sound, but was curious about what kind of fire breathing beast was required.
Try auto rate, no need to take 44.1 all the way to 48 x 512. Just extra computational power needed to do it. I think auto rate ends up sounding better as well. Or that might be in my head.
In double precision computations, the Titan line (like the Quadro work station lines) are greater than double the performance of other consumer graphics cards like the 980, 1080, etc. Why? Because NVIDIA clips/cripples double precision computation performance of the consumer lines to prevent eating into sales of the Quadro line. This is also why there is a higher premium on the Titan.
I am pretty sure that during the filters initialization the GPU is not involved, just the CPU. Then once playback starts the GPU is now active. I have a 6700K processor with a 1060 card and I cannot do the non-2s XTR filters at 512, the 2s are fine as is every other filter except the non-2s xtr.
Only thing I have that can do non-2s xtr filter to DSD512 is i7-6950X CPU with 4x DDR4-2400 CL12 RAM… And then it is already running all 10 cores around 90% load…
I want to see how the new Ryzen 9 with 16 cores will do and what will be the price…
Can you tell us some more about the xtr filters, where they fit in terms of length, phase change, bandstop, frequency and time domain, pre and post ringing etc. Why are they so computationally intensive ?
I created it for people who like Chord Dave -style filters. It is about 5 times longer than poly-sinc and thus about 5x heavier. Offered in linear- and minimum-phase variants. Stop-band attenuation exceeds needs of 40-bit PCM, being over -240 dB. Normal poly-sinc filters are created to reach excess of 32-bit PCM resolution (> -192 dB) stop-band attenuation.
Since high-rate modulators can reach peak SNR higher than 32-bit PCM resolution at audio band frequencies, I thought it could be useful…
Coincidentally, the June issue of Stereophile just published a review of the Chord DAVE. In the digital filter measurements, check out the length of that FIR ringing, symmetrical both pre and post. The MQA advocates must be rolling in their graves.
