Recently I've had the whim to play through Final Fantasy 6 (3 in the US) again, and have been doing so on an (SNES) emulator. I'm currently about to start the last level, but that's beside the point of this post. While playing, I thought I'd take a look at the SNES hardware and write a blog post about it, given that I'd already looked at the NES hardware - though not anything half so extensive as what I did with the NES, just a look at the basic hardware. What I found actually surprised me. As it turns out, the SNES really is the Super NES; that is, it has very similar hardware, only better.
First of all, the CPU is a generational upgrade to the NES' CPU, bigger and better. The CPU is now 16-bit (as opposed to the NES' 8-bit CPU), but has essentially the same instruction set (with some augmentations). The CPU is still CISC, sporting the same three general registers (an accumulator and two index registers) and operating on an accumulator register using the contents of memory. However, the CPU now has much greater flexibility in memory access, with a 24-bit pointer register and the ability to access memory relative to the stack pointer*. It also now has multiply and divide instructions, as well as a few other things. But deep down, it's the very same instruction set architecture, and in fact has an 8-bit compatibility mode that lets it directly execute code from the original NES CPU.
*Both of these features were conspicuous absent in the NES, as I believe I noted. As the NES only had 8-bit registers, it had no way to hold pointers (which were 16 bits) in registers; to make use of pointers the NES had an indirect addressing mode where the CPU would write a pointer to memory 8 bits at a time, and then had an instruction to load/store a value through that pointer (think "mov reg, [memory]"). While the NES did have a stack, it had only push and pop instructions, and lacked the ability to access data relative to the stack pointer, preventing use of the stack to pass parameters or store local variables; consequently, parameters and local variables were assigned fixed memory addresses, and the stack was rarely used.
The situation is somewhat similar in the graphics system. While the graphics chip is drastically more powerful than the NES', it's based on the same concept of background layers and sprites, all drawn from a (larger) bank of 8x8 tiles. The SNES supports twice as many sprites as the NES (and a lot more per line) and sprites can be much larger (up to 64x64, compared to 8x16 on the NES), with 16 colors each (compared to 4, including transparent, on the NES); but perhaps the most interesting improvement is that there are now 4 background layers, and they can be combined via various raster operations in many interesting ways.
To illustrate this, take a look at this picture: a typical shot from a battle. The SNES supports 8 different graphics modes. For the most part the difference between the 8 is the number of layers and how many colors each layer supports (presumably this is due to it being too expensive to put enough video memory in to allow full color from all layers at once). In this scene I'm guessing that it's using mode 3 ("3 layers, two using 16-color palettes and one using 4-color palettes"), based on the number of layers I can see plus the number of colors.
To show off the graphic abilities of the SNES, next is a screen shot of the special effects from casting of a spell, which we're going to dissect.
This is layer 1. You can see the bubble from the spell here, sporting transparency that tints other layers. Interestingly, you can see part of a screen (the magic selection screen) that isn't open at all; in the composite, this section is covered up by the bottom part of layer 2. In other words, in the top part of the screen layer 1 is drawn on top of all other layers, while the bottom part is drawn at the bottom of the layers; this goes to illustrate what I said about very complicated and flexible interaction between the layers (it's possible this involves changing the layer parameters in between lines, a technique I mentioned being possible on the NES).
Layer 2 is the background for the whole screen, the top section the battlefield, the bottom section the menu system. Note that a sine wave offset pattern has been applied to the battlefield background; while I haven't investigated it to be certain, I suspect this is accomplished by simply modifying the screen scroll position in between drawing each line.
Layer 3, again serving an array of purposes, is used for special effects and the text on the menus. You can't see it clearly at all from just this layer, but this layer is used to produce those discolored blotches on the spell bubble. This may be another case of transparency for the special effect, but it's hard to tell from screen shots alone.
To illustrate this fact, layers 1 and 3 together:
And finally, the least interesting part: the sprite layer. Although here again we see something rather unexpected: it looks like there's some type of sprite garbage that is normally (again) covered by the menu.
So, that's basically what 7 of the 8 graphics modes are all about. The 8th one, known as Mode 7, however, is a bit different. It has only a single large 256-color background layer, but it has the ability to apply a transformation matrix to the background, allowing scaling and rotation. This is used very commonly in SNES games, especially with the swap-the-registers-between-lines trick, allowing it to do primitive 3D perspective projection. Believe it or not, that minimap is actually a sprite (as is some of the glow in the background).
One particularly interesting (in the sense of peculiar) feature of the SNES design is the sound system. As opposed to most sound systems, the SNES system does not consist of the CPU simply writing channel parameters such as sample #, frequency, etc. to the sound chip which then performs the requested operation. Rather, the SNES has a separate CPU (the SPU) which acts as a sound coprocessor: sound programs are written, assembled, and then executed on this coprocessor; once the program has been uploaded, the sound system can play (e.g. music) without any further involvement of the main CPU at all (I proved this a decade ago by showing that shorting between two pins on the SNES would crash the main CPU while the SPU continued to play the music without missing a beat), though obviously the main CPU must instruct the SPU when it's time to play dynamic sound effects.
We can only guess why the SNES was designed this way. The most obvious possibility is that this frees the main CPU from having to deal with music and sounds effects, leaving it more cycles to spend on something else (especially in the case where one or more channels of the music must be temporarily dropped to allow a sound effect to be played).
An alternate possibility that I haven't been able to confirm is that this is done to increase the resolution of the audio system. In Blaster Master, the game would perform all the calculations for the frame, then spin waiting for the vertical blank interrupt to begin work on the next frame. Thus code executed 60 times a second, apparently including audio code. If this is true in the general case, that limits the resolution of audio operations to 1/60 second as well. In contrast, the SPU runs at 1 MHz independent of the main CPU, allowing it to issue commands to the sound generator at any time, in theory allowing for higher music tempo and more complex audial effects.