Post
When I tackled the question of rendering webpages visually... It struck me that for the uses I had in mind (document & GUI rendering) existing & historical GPUs were either significantly underpowered or overpowered. So it interested me to design something which was just right!
For one I designed a "Compositing Coprocessor" which iterates over "sprites" sorted by top-left corner & per-pixel performs 10 vectorized multiply-sums or texture-fetches as it outputs a DVI videosignal.
1/3?
This design allows our (second) GPU to render an arbitrary number of sprites (say, 1 for every letter) whilst being flexible enough to handle all of CSS's visual properties! Usually in a single rendering pass.
The caveat is that for this to be realistic we'd need those sprites to be sorted, where we could take advantage of layout constraints to take shortcuts.
But given how remarkably weak our CPU would be at math, I had to design another GPU to calculate onscreen layout.
2/3!
A "Layout Coprocessor" would perform simple arithmetic over concurrent tree traversals. Internally these would be binary trees, but a layer of abstraction would allow more flexibility.
A mix of physical connections between compute units (which would double to facilitate running sums & multiply) & shift registers would facilitate traversing the tree. There'd be a measly 256 of these Compute Units controlled by their choice of 4 Control Units.
Loosely esembles modern GPUs but lower-spec.
3/3Fin
This morning (see above) I summarized some GPUs I designed to augment my hypothetical "string-centric" CPU. So what value do they add?
Basically 2 things:
1. Greater output bandwidth.
2. Vectorized 32bit arithmetic.
I had great fun exploring how little computers really need to calculate, but... Computer graphics does need it! At least once you get beyond 8bit era graphics.
Install bonfire.cafe
Get the full app experience
