Bruno Putzeys Talks DACs
Eureka! I remembered our conversation at RMAF! Kidding. Philip O'Hanlon was not only smart enough to record Bruno explaining aspects of his approach to DAC design to me, he was kind enough to allow me to share it here.
Here's Bruno Putzeys on his Mola Mola DAC:
So we're looking at a discrete DAC although these days I think that you can have very respectable DAC chips. They all still have shortcomings which I felt had to be resolved. For example, one of the major shortcomings you have with a classical DAC chip is that the noise floor is signal dependent. Whereas analog noise is just a steady hiss that's always there at the same level, these DAC chips typically have a noise floor where you have spikes popping up depending on what the signal is.I mumbled, "Absolutely, yea, yea." Philip O'Hanlon added, "Now what's interesting is Bruno said to me yesterday have you listened to the DAC with a regular CD file. I don't know if we played any for you but before we go over anything else let me do that. Let me play a well recorded piece of Redbook and the interesting thing is that with Bruno's topology he is upsampling this and I would say to you it's the only time I've ever heard a digital file upsampled where it sounds outstanding. I mean everything else makes it sound different but it doesn't make it sound better. That would be my ten cents worth."
Also linearity at low and high levels is an important factor for me. I've found that if you want to make a converter where the output signal really sounds the same as the input converter, and having made an A to D converter I can obviously do that test, you really need staggeringly low distortion figures at all these signal levels. That's why for Mola Mola I've decided to base the design on a pulse width modulator (PWM), that I designed in 2004, that has enormously good linearity.
That PWM then drives a string of discrete 1-bit DACs which are also time delayed by a single clock cycle. The clock frequency of the DAC is 100MHz and there are 32 single-bit DACs arranged as a shift register. All of these outputs are then added together electrically and converted into voltage.
The modulators run on DSPs. There are three DSPs in the converter. The first DSP is responsible for upsampling any input signal to 32-bits at 3.125MHz. Then each DAC channel has its own DSP that converts that into PWM with a clock rate of 100MHz and a sample rate 3.125MHz. Also included in that first DSP is a brand new asynchronous sample rate converter or de-jittering algorithm which will pretty much remove any jitter that it might encounter.
We have done tests with different input cables and what have you on the digital side and nothing happened, it always sounded the same, and I think you've heard it yourself—it sounds spectacularly analog.
Back to Bruno:
Well yes, the upsampling is of course a kind of a standard part of most D to A converters but how you do that upsampling, that is how you treat the...how exactly you work out what the intermediate samples are, there are a number of shortcuts that practically anyone makes which are absolutely deadly to the sound. What you typically find is that the DAC is supposedly perfectly flat but if you zoom in microscopically then you find these tiny ripples, like a hundredth of a decibel, which repeat themselves all over the audio band and which corresponds to a small pre-echo and a small post-echo. Not just the ringing but also the echo. And that completely kills all body in the sound.Me again, "Fascinating. What you said before, thin, I think you described the sound as 'thin'", Bruno interjects, "As lacking in body", "Yea lacking in body, I find that to be the case." (Gosh I really did not have much of value to say, did I?)
The reason why they do this is to save on computing power. Another classicical short cut is making a fairly sharp filter like the classical brick wall filter which is then centered around exactly half the sampling rate which means that you still get some aliasing which is not as evil as you would think but which is just kind of a mathematical gimmick again to reduce the number of cycles and the amount of processing power that you need. But it's not necessarily psychoacoustically optimum. Psychoacoustic optimum actually starts rolling off a bit earlier so I'm actually sort of getting the perfect specifications amiss when we're playing Redbook. It's not flat to 0dB at 20kHz it's already 0.1 of a decibel down at 20kHz which is tiny but it allows you to reduce the amount of ringing on the impulse response. Also as I said a tiny bit of aliasing is not a problem sonically it's not that dramatic so actually allow the transition band to be a bit wider again to make it a bit shorter.
In listening tests I just found that the classical transition band, which is literally 10% of the Nyquist bandwidth, is just too narrow. It's just enough to be audible. Make it a tiny bit wider and immediately the signature of the brick wall disappears.
"I mean there's, there's a lot of, um, the way I say it," this is me muttering again, "is there's lots of digital out there that sounds like OK digital, then better digital, but it hardly gets beyond sounding like really good digital." (sometimes I get thoughts stuck in my head and I'll use any excuse to get them out).
Yea, and ideally of course you want to use a similar design procedure on the recording side but of course we don't have any power over that. But even by just doing it correctly on the replay end, you're already halving the problem. And that really makes a significant difference even with quite garden variety Redbook recordings.I think I'll bring my digital recorder to every hi-fi show.