The 8250 (the equivalent disk drive for the PET) had two 6502s inside!
In today's language, one was the application processor running DOS and the other was a DSP that did the low level hardware monkeying; they were connected up to a shared data bus. Because the 6502 only does bus accesses on one phase of the clock, by inverting the clock for one processor, it was safe to connect them both up to the bus simultaneously with no loss of speed...
The really fun thing about these was that you could actually download code to them and run it, so if you did have code that was small enough and CPU intensive enough and required little enough IO, you could in fact use them as an accelerator of sorts (but limited by a slow serial port and 2KB(?) I think of RAM on the drive)
I know of various code that downloaded code to it, but usually loaders or just to do things like blink the leds - I've never seen code that actually used the extra CPU for anything serious. I'm really curious if any exists.
It worked by transferring the "live" GCR-encoded data from the 1541's disk head to the C64 and simultaneously doing a fast checksum. Part of the checksumming was done on the 1541, part was done on the C64. There simply weren't enough cycles left on either side! Most of the transfer happened asynchronously by adjusting for the slightly different CPU frequencies and with only a minimum number of handshakes. This meant meticulous cycle counting and use of some odd tricks.
A GPU isn't really comparable to the modern CPU sitting beside it, since it lacks many of the CPUs features, eg. virtual addressing, while in the other instances mentioned the processors were pretty much the same, feature-wise.
GPUs do, like most peripherals, contain additional CPUs that control them -- in the case of GPUs these are proprietary architectures specifically designed just for scheduling and sequencing tasks on the GPU and are said to be quite potent, for a device control processor.
Similarly hard drives had multi-core MIPS and ARM processors for quite a while (ditto for SSDs, in both cases they are ASICs, the actual data juggling isn't done by the CPU cores), and they even have quite a bit of RAM -- the cache -- that they'd normally share with the DMA engine, plus some of their own...
The control processors in most other peripherals are usually much weaker, though. They're pretty much just microcontrollers.
