Oversampling, often combined with noise shaping, attenuates quantization
noise. It does not improve linearity.

If the core DAC is one-bit in resolution, it is infinitely linear,
ignoring minor secondary effects. But it would take too much oversampling
to achieve 20- to 24-bit resolution.

If a multi-bit DAC is the core, a practical oversampling ratio is enough.
But the linearity of the core DAC becomes an issue. Usually the core DAC
is built with unit elements (still not matched), and the elements are used
randomly (hence scarmbling, etc.). It effectively converts nonlinearity
(distortion) into noise.

Stewart Pinkerton wrote in message news:g2XEc.8215$MB3.5340@attbi_s04...
On 1 Jul 2004 00:57:51 GMT, (Marcus) wrote:

I suspect it is impossible for the components to be matched
better than maybe 20-bit precision.

Which is why no one tries............

I looked up the dCS converter and I can see it has some technology
to cope with this problem. It is also known as scrambling,
randomization, or averaging, which is similar to what the IC
foundries do.

No, it's known as oversampling.

AFAIK, all 20- to 24-bit DACs (monolithic or
discrete) are based on a multi-bit core DAC with randomization,
and of course oversampling is in place. In that sense, they are
quite close in architecture.

All '24/192' DACs (including the dCS units) are in fact between 1 and
5 bits in basic resolution, and gain their 24-bit rating by high
oversampling. This is not a new technique, as the very first
Philips/Marantz CD players back in 1983 used a 4x oversampling 14-bit
DAC to achieve 16 bit resolution.