In the PSX/PS2 era (or really, the era starting from the SNES's SuperFX chip), the dedicated graphics ASICs in consoles, combined with the fact that those ASICs were being targeted individually at a low-level by game devs, were putting out results that seriously outpaced what you'd expect out of your PC's 3DFX Voodoo card.

That wasn't because the designs were more clever, mind you; but just because the hardware designers didn't need to think in terms of an architecture that contained concepts like dynamic frequency scaling and multi-monitor support and a kernel that blocks on disk IO. Consoles were hard real-time embedded systems, and the games were their unikernels; well into the PS2 era, console game were still relying on VBlank interrupts for physics timing!

And what this got you, was effects that were only achievable on an $8000 SGI workstation, for $300. Slightly-beyond-state-of-the-art, for cheap. But in exchange, it forced heavy consolidation in the console manufacturer market, because developing that specialized hardware wasn't cheap (like it was back in the 8-bit micro era.)

But "generic" PC GPUs eventually started scaling in power geometrically, to the point where the specialization and hard real-time guarantees just weren't needed any more to achieve modern graphics cheaply. The low-level-targeted specialized-graphics-ASIC technique wouldn't be of much benefit today, because six months later there'd be another new "generic" GPU out that could do what that ASIC does without breaking a sweat.

The same thing happened in networking: ASIC switches with special RTOSes were needed to run data centers—until CPUs and PCIe advanced enough to take over. Now everything (except Internet-backbone gear) is Software-Defined Networking, i.e. generic boxen running VMWare EXS running a Linux/BSD router VM.