It is well known that the PlayStation 2 implemented compatibility with the previous console by essentially incorporating a PS1 on a chip.
The fact of backward compatibility is unremarkable as far as it goes. Many computational devices over the years have done that, indeed the vast majority of all programmable computational devices ever manufactured have done so, for the obvious reason that the ability to run existing software is a valuable feature that attracts customers.
But the chosen method is actually rather extraordinary when you stop and think about it. The IBM PC AT did not incorporate an original PC on a chip. Nor did the BBC Master thus implement compatibility with the BBC model B, nor the Macintosh 2 or the iPhone 2 or the iPad 2. They all provided backward compatibility by being supersets of the previous version. To the extent that they contained duplicates of the circuitry from the previous version, those duplicates were dispersed in the new circuitry, and used for their intended purpose even when the new machine was being used in its new mode. For example, when a 386 is running 32-bit code, the old AX register is part of the new EAX register.
But the PlayStation 2 implemented backward compatibility by segregating the entire previous machine into a separate chip (which, when a PS2 game was running, was used, if at all, only for some secondary purpose). Why?
It's not about the CPU instruction set. Both machines are based on a MIPS CPU. The PS2 CPU also provides a high-performance vector unit, but that's as well as, not instead of, the MIPS core.
Is it to do with the graphics chip?
But generations of PC graphics cards have provided backward compatibility, and as far as I know, they have not done so by segregating the previous GPU onto a separate chip.
The only salient difference I can think of is that PC graphics cards are accessed by device drivers. Could it be that there is something about GPU instruction sets that makes it unreasonably difficult to provide backward compatibility by the superset method normally used by CPUs, such that easy backward compatibility can only happen when games are screened from the hardware by a device driver, and when games have directly accessed the hardware like on a console, there is nothing for it but to segregate the old GPU onto a separate chip?
If that is the case, then why are GPUs so different from CPUs in that regard? And if not, then what is the explanation?