Many lanes. One kernel.
Schedule integer matrix operations across 4, 8 or 16 logical lanes. Follow every multiply and add.
SIMT-INSPIRED EXECUTION ↗A GPU is an architecture before it is a chip.
What if its instructions, memory and execution
could travel together — onchain?
WG–01Instructions are only half the machine. We bring lanes, banked memory and a replayable trace into the same experiment.
Schedule integer matrix operations across 4, 8 or 16 logical lanes. Follow every multiply and add.
SIMT-INSPIRED EXECUTION ↗Route reads across 2, 4 or 8 memory banks. Discover why more lanes can mean more contention.
HBM-INSPIRED BANKING ↗Export the inputs, configuration and trace as a hashed artifact. Import it and recompute the result.
DETERMINISTIC LOCAL REPLAY ↗A real computation. A deliberately tiny machine.
Local architecture simulator · no blockchain transaction
> WG–01 initialized. Waiting for kernel.
SHA-256 identifies the artifact. It is not a validity proof or a blockchain receipt.
AI made compute physical again. More transistors. Taller memory stacks. Bigger bandwidth budgets.
Wafer Goblin asks a different question: what is the smallest complete machine we can describe, execute and independently replay?
Our direction is an inscribed GPU specification: instruction semantics, memory rules and reproducible execution. This workbench explores that idea with integer kernels in your browser. It does not reproduce physical HBM bandwidth, run AI models or execute onchain.
Start small. Inspect everything. ↗