From another thread:
About Chord and upsampling and M-Scaler…
When Chord released the M-Scaler some people (including me) were distracted by the term “upscaling”. The $5,000 M-Scaler upscales 44k to 705k and feeds that into Chord’s DACs, but Roon already does upscaling (for $0 including VAT) so why do I want the M-Scaler? And much as I admire Rob Watts’s work, and his openness in lectures and writings, I fear his explanations in this area drown us in detail and are not helpful.
I read up on it, stubbornly, and I think I understand (and I eventually bought one). And with characteristic hubris, I think I can explain what it does better than Rob can. Sometimes knowing less helps…
A few years ago, people used to say that DACs perfectly recreate the original sound because Nyquist said so. If you believe that, I have a 1982 Sony CD player I can sell you for $20,000, a bargain price for perfect sound, no? “Ok, those early devices were flawed but today DACs are perfect because Nyquist said so.” Where in Nyquist’s paper does he mention 2019? “But the paper says perfect recreation, are you saying Nyquist is wrong, do you claim to know better than Nyquist?”
Poor Nyquist never said anything of the sort. Nyquist and Shannon did math, not engineering, they never claimed that their solutions are physically realizable. “But math is the foundation for engineering!” Yes, but math does not have a problem with solutions that involve infinite numbers. Infinite is awkward in engineering, because the cost is infinite, and energy consumption, and latency. And size and weight of the equipment, and there is the black hole problem, and destroying the universe. And note that Moore’s law doesn’t help with infinite.
“But we can get close to Nyquist’s perfect.” Maybe, but like most math results, Nyquist’s theorem does not address how close to the solution you get with an incomplete implementation. That’s an engineering question,
This is what Rob Watts addressed. Looking at the infinite sinc function that is used in Nyquist’s perfection, he figured out how much computation he needed to reduce the residual error below a specific threshold. When he first did this, the requisite computation wasn’t feasible, so they made the best implementation they could, and gradually improved it, and they now claim the M-Scaler has achieved that goal. So this is not just the usual incremental improvement that the industry (including Chord) does: “this year’s model is better than last year’s model”. Specifically, he says the error (distortion) is below 16 bit, which is 96 dB or 0.0016%. This is where the million taps comes in: the previous devices have increased the tap length as they could afford to, but the million taps threshold was always there, given the target error threshold of 16 bits.
And as this approximation of Nyquist’s perfection within a specific error margin was really the objective, I believe the upsampling is incidental. The calculation delivers its results to the DAC in upscaled form, because that is the way the calculation engine can cooperate with the DAC. Why? How does this two-stage processing work? I don’t know—there is a lot of discussion on the internet if you care, but I consider that secondary. Why is the million-tap scaler a separate box and not built into the DAC? I don’t know, business reasons?
Anyway, with this understanding, we see that the M-Scaler doesn’t necessarily have anything to do with Roon’s upsampling, or upsampling in any other device (my Meridian digital speakers upsample to 705k). What is central is the algorithm, how it does the calculation to bring the error down.
(How fabulous is it? This is a technology discussion, not a product review, I’m not going to gush over how it was like removing a veil and how it opened up the sound stage and how it was just like vinyl and how the backgrounds were blacker and how the cymbals shimmered…)