No frame buffer. The target CRT display scanned an electron beam from left to right in about 63 microseconds, snapped it back in just a couple more microseconds, working gradually downward to produce about 200 such lines before taking a couple milliseconds to return to the top. The 6507 had to literally write the pixels into that raster in real time, and was just skin of the teeth able to do that, hence the phrase 'racing the beam', leaving just the retrace times for any other program code to run. To get video from a cart, all that still has to happen. Pretty cool.

It reminds me of the gameboy player for the SNES, where the entire gb console ran inside the cartridge and used the snes for inputs and display. In this case a PIC controller is the cpu in the cartridge doing the heavy lifting.