Mode 7 on the SNES is not 3D graphics. It is one matrix per scanline

The Super Nintendo's graphics processor supports eight background modes. The seventh differs from the rest in giving you a single background layer — a large one, 1024 by 1024 pixels, in 256 colours — but running the coordinates through a two by two matrix before output, set by four registers, with its own pivot point and offsets.

A two by two matrix does exactly three things: rotation, scaling and shear. This is called an affine transformation, and there is no perspective in it. Mode 7 by itself does not draw a three-dimensional space. It draws a flat picture that can be turned and zoomed.

Where the road comes from

The perspective comes from somewhere else, and that is the interesting part. The matrix registers can be changed not once per frame but per scanline, through HDMA — direct memory access tied to the horizontal sweep.

The arithmetic that follows is simple. The upper rows of the screen are what is far away, so they need a heavily compressing scale: a large piece of texture fits into one line. The lower rows are what is near, so the scale is large. Recompute the coefficients for each of the frame's two hundred and twenty-four lines and the plane recedes toward the horizon and becomes a road.

The three dimensions are not in the chip. They are in the fact that every line uses a different matrix, and somebody worked out in advance which one.

What it costs

The single background layer is spent on the floor. That means there is no real backdrop left and everything else has to be sprites: rivals, barriers, interface, sky. In F-Zero, released on 21 November 1990 alongside the console itself, the cars do not rotate — they are pre-drawn sprite frames swapped in by heading. The sense of volume is assembled from a flat road and flat pictures above it.

Super Mario Kart in 1992 added a second difficulty to the same technique: a split screen. Two halves require two independent sets of coefficients plus room for the interface, and one central processor was no longer enough for that arithmetic. So a separate DSP-1 chip, computing the perspective transforms, went into the cartridge.

That is an important detail about SNES architecture in general: the console got its serious computation from the cartridge. The DSP-1 was followed by the Super FX in Star Fox in 1993 — effectively a second processor plugged into the slot along with the game.

Why remember this

Mode 7 is a model case of an effect that is ninety per cent programmer discipline and ten per cent hardware feature, but entered history as a hardware feature. The Mega Drive had no such mode, and developers did the same thing in software, paying in processor time: the road in the OutRun ports and Sonic's special stages were assembled by hand.

Today's vocabulary works the same way. Hardware ray tracing, mesh shaders, neural upscaling: these are primitives. What comes of them depends on who calls them and how, every frame. Marketing sells the primitive, and what we actually see is always somebody's work on top of it.