work

GPU and streaming rasterizer

The active development track within SPU-13 is a graphics pipeline built on the same exact-arithmetic foundations as the rest of the platform.

Its distinguishing property is what it leaves out: there is no framebuffer. Pixels are generated in step with the display beam rather than composed in memory and scanned out afterwards, so the design uses no external memory bandwidth at all.

Where it has reached

Three results on physical silicon in three consecutive days, each with a dated ledger entry, a pinned bitstream hash, and a prediction registered before the hardware was programmed:

• First video output — 640×480 at 60 Hz, 0.006% timing error
• First rasterized geometry — a triangle, measured against its design rather than merely observed
• First depth-resolved image — two overlapping triangles resolved per pixel by interpolated depth

See Evidence for the measurements and photographs.

What it is not, yet

There is no 3D, no projection and no transform stage; the geometry is hand-computed constants compiled into the bitstream. There is no shading, no texturing and no antialiasing. There is no host link, so scenes are static and changing one means rebuilding.

Those are the next pieces of work, in roughly that order. They are not claimed as present.

Why the constraints matter

The whole flow is open — Yosys and nextpnr through openXC7, no vendor tooling. That toolchain cannot place differential outputs, which rules out HDMI and is why the display path is analogue VGA driven through a three-resistor DAC.

Working inside those limits is the point rather than a compromise: it keeps the entire design reproducible by anyone with the same open tools and the same board.