Why so many op amps or power amp IC's use asymmetric NPN output?

[ELS122]

For example the LM1875. uses asymmetric NPN output.
And doesn't even have a baxandall diode... how did they get such good THD specs with an amp layout like that? And why NPN asymmetric output in the first place?

[Kleinstein]

In the old days there were no good PNP power transistors.  So the asymmetric output stage was used as a work around. Even today the PNP silicon transistors tend to be a bit slower because of the matrial properties.

For the integrated chips using a cheaper process the quality of PNP transistors can be quite bad. So for ICs it is still attractive to prefer NPN over PNPs. There are also manufacturing processes to make good PNPs, but these tend to be more expensive.

There is not much advantage from symmetry in the output stage. The main way to get low THD is having enough and fast enough feedback

[magic]

So-called complementary bipolar processes, capable of fabricating similar NPN and PNP on one die, gradually evolved from "horrendously expensive" in the 20th century to merely "expensive" in the 21st. These days you can get some LM4562 or OPA1602 for a bearable price, but old parts were designed (and are still made) on basic, not-so-complementary processes with large and sluggish PNP.

These processes have also evolved, from μA709 which could only use PNP in common base mode, through the μA741 era when they made a viable output stage, to RC4558 where a decent input stage became possible and later variants like NJM2043/NJM2068 with surprisingly low noise and a few MHz bandwidth.

Out of old parts (edit: with non-trivial output power), OP27 and LT1028 come to mind as rare examples of decent opamps with complementary (diamond buffer) output stages. NJM4580 has single emitter followers, is often found in low-end audio gear. On the other hand, NE5532 uses almost only NPN in the signal path, and the PNP second stage is unceremoniously bypassed capacitively above hundreds of kHz.

There is not much advantage from symmetry in the output stage. The main way to get low THD is having enough and fast enough feedback

In this kind of topology you never have "enough feedback" at high frequencies, because AC open loop gain of the input stage and VAS is reduced to maintain stability. Meanwhile whatever local feedback exists in the output stage is itself "distorted" by modulation of output stage loop gain with load current.

[CosteC]

Silicon technology has its limitations and its advantages too. Pairing discrete transistors is manufacturing nightmare, and those will thermally drift still as thermal coupling is poor.

Inside IC having two paired transistors, well thermal coupled is easy. Adding 5 more transistors for compensation is also easy, but those are usually not put on datasheet schematic as not needed for understanding how it works.
Making transistors with accurate current densities ratios is also easy in IC and impossible in discrete constructions.

Somebody with experience in silicon technology can explain why mixing NPN and PNP in one structure is not preferred in some cases.

[Kleinstein]

With a low number of mask and diffusion steps there are just no vertical PNPs and only lateral PNPs as a makeshift workaround. The lateral PNPs, especially with low resolution masks have pretty poor performance (e.g. gain of 5 and low speed). To get vertical (more classical PNPs) it needs another (maybe 2) mask and diffusion step and this makes the chips more expensive. The extra diffusion step also stresses the parts produced before and need more (especially more parameters) fine tuning of the process steps. So it is more than adding just the few extra production steps (e.g. 7 instead of 5 masks).

Another point is that the NPNs are still usually a bit faster than PNPs. So one may get higher speed with a NPN only design.

[David Hess]

Complementary bipolar processes are more expensive than an NPN only processes, and I suspect this is even more so when high power PNPs are required.