close to metal
Understand the hardware first. Abstractions earn their place.
A small open-source lab for understanding and pushing accelerator hardware—from tensor algebra to kernels to distributed execution.
compute/compile/accelerate
01 / principles
Understand the hardware first. Abstractions earn their place.
Performance claims should collapse into numbers, traces, and reproducible tests.
Prefer sharp primitives over sprawling frameworks and hidden machinery.
Source, experiments, failures, and useful discoveries belong in the open.
02 / from tensor to silicon