I think full 3d audio is a different problem though because it, at least, requires a version of the rendering problem (for waves). It's harder in some ways than light rendering (because phase/coherence matters sometimes, the wave equation is harder to solve), but easier in others (no need as much detail as light, wavelength is large), or just plain weird (nonlinear effects from rattling and such).
Or there’s still some “generic” box (I mean that you can program yourself) like the Symbolic Sound Kyma Capybara. They’re quite niche thought, like modular synth.
There are always slightly harder, and better ways of doing something which the accelerators are better at. Audio acceleration I guess peaked too early or they just couldn't get the tech demos as impressive as graphics.
I remember reading about a demo, which I believe was from Matrox. They managed to get an audio 3D environment working over a pair of stereo headphones which was good enough that you could play Doom headless. Just close your eyes and you could tell where people were.
A lot of what I read about here is that prerecorded samples is good enough, in the same way that raster lighting is good enough and raytracing is a waste of time.
It is cheap enough that it is bundled with everything. But that doesn't make it better than a discrete GPU/Audio chip.
Intel GPUs are barely good enough to run an OS desktop yet they hold the majority of the market.
To be fair, realtime synthesis just became obsolete for most purposes once CD quality digitized audio became cheap enough to store (and later, to stream). And for musicians, once CPUs became fast enough, SW synthesis with its limitless possibilities took over from HW synthesis.