t’s because they want the passband to be flat to 20.0 kHz @ 44.1 kHz Fs and can’t / won’t design a filter with enough taps to be into the stopband by Fs/2. If you notice, almost all are -3 dB at 0.49*Fs. Passband is almost always flat to 0.454*Fs which is 20.021 kHz @ 44.1.
I don’t know which manufacturer started it, but ALL of them do it, starting way back in the day with TI and Analog Devices.
There’s really no issue here on the DAC side. On the ADC side, there can be aliasing as a result.
If memory serves these are called half-band filters, and there must be substantial computational advantage associated with this approach for it to be as popular as it is. Last time I saw the additional choice of a "correct" filter, that had a fair bit more periodic passband ripple than the "normal" ones. If you can get better-behaved in-band characteristics and/or lower latency for a bit of aliasing up top where nothing much tends to happen anyway, that seems pretty compelling.
BTW, maybe check out the CS5396 vs. CS5397 datasheet - these were basically the same chip but one with a "normal" filter (for audio) and the other with a "correct" one (for measurement).
Some here should think harder about why it matters that we be in the stopband at exactly Fs/2 of the original sample rate when the destination (interpolated) rate is 8x that. The only thing you gain is possibly chopping off some garbage if the ADC had a poor filter in it.
This I think was what Bruno Putzeys has referred to as the "burnt steak" approach. I'll have to dig out that presentation again.
If you happen to be overly fussed by these issues, they're nothing that a bit of oversampling and software SRC couldn't fix.
BTW, Crystal (later Cirrus Logic) had one of the first delta-sigma DACs with effective switched capacitor lowpass filtering in about 1993-ish (CS4328 I think it was). Philips had used some before in their SAA7350 - TDA1547 platform but it was only partially switched capacitor and partially conventional, so effectiveness in terms of jitter suppression wasn't as great. There was a discussion of this in
The Audio Critic back in the day.
I think for a while up to the mid-'90s or so Crystal and AKM had some sort of cross-licensing agreement or other IP sharing, at least their CS5389 and CS5390 era ADCs seem to have been pretty much interchangeable. Their paths seem to have split after this, with AKM releasing the AK5391 in '96 and their last 48-kHz-only flagship ADC (AK5392) in '97 while just a few weeks earlier that same year Crystal had already made the jump to 96 kHz (CS5396). Must have been a bummer for AKM. Not sure why they were putting off the transition to double speed so long anyway, the writing was definitely on the wall by this point.
There is relatively little doubt that in the late '90s the CS5396 was about the best audio ADC chip you could buy (Benchmark used it at the time). They never really won that crown back after being overtaken by the AK5394A, which continued to define SOTA ADC chip performance for over a decade, and continued to play a good second fiddle. (The CS5381 is "close but no cigar" terrain but still a good part in its own right. It's what the Behringer 2496s have switched to after the old AK5393 was retired, nice upgrade there.) There was very little wrong with the CS4398 DAC either, it even has some headroom for intersample-overs. One peculiarity of theirs was using partially IIR (minimal phase) filters when everyone also was all FIR (linear phase). I think this has gotten way more common now.