they're really simple, in their basic form and for simple ISAs!
it mostly boils down to keeping a bunch of registers and a giant switch statement. Each case simply implements the opcode. You have an array of bytes for the memory, and some emulated devices (e.g. trigger a screen update when the framebuffer memory gets changed, or set the instruction pointer to a handler when a key gets pressed.)
It gets hard when instructions need lots of decoding, or you have 3d graphics hardware to emulate, or if you have something like a BIOS, or if you want to JIT.
I'm actually eyeing implementing a console on an FPGA as my holiday project, with something like Chisel.
It seems like a lot of the complexity may also come in when you have important circuits to emulate other than the CPU. In that case you'd have more to worry about in terms of timing, dataflow, and synchronization. And some platforms might conceivably have analog circuits that play an important role too, although I guess you might be able to abstract that out by trying to create a rough digital functional equivalent even if it isn't completely faithful to the behavior of the analog part.
that's true! but I still think you can get a lot of the way with a NeoGeo emulator without it being perfect on timing. that's a lot of the reason emulators work well on most games but fail to replicate the right graphics or sound in some - some games played hardware tricks to optimize.
My friend who works on the MAME project told me that it's a very friendly platform for adding new emulators (whether you want to add an emulation of a game, a device, a platform, or whatever) because it provides a good structure to work with and a lot of tools to describe recurring patterns in electronics and systems. That might be a good middle ground to start with -- trying to add support in MAME for some device that's not presently emulated there. (Apparently the MAME project really appreciates such contributions, since they think of themselves as pursuing mainly a historical preservation and documentation mission, and would always like to see more devices covered.)
Although that's not "from scratch" because it would still be using their libraries and plumbing, it could be "from scratch" in the sense that you might take an environment with 0% support for emulating a certain device or system and build it up to having 100% support. So it seems like a nice way to start.
A long time ago I wanted to code a neogeo emulator but gave up before I even started, I didn't have a clue where to begin.
I am amazed at anyone that can code an emulator from scratch.