/ start-here
00 Start Here
The mental models. Why kernels decide who wins, how to think in speed-of-light terms, and the exact skill map a GPU kernel engineer is hired for.
→
Why kernels run the world
Every frontier model is, at the bottom, a pile of hand-tuned kernels. Here is why that layer is the one worth learning.
→
The three regimes: compute, memory, overheadBRRR
The single most useful mental model in performance: every kernel is bound by compute, memory bandwidth, or overhead. Learn to tell which.
→
Speed-of-light thinking & the rooflineROOFLINE
Arithmetic intensity, peak FLOPs vs peak bytes, and the roofline that tells you the fastest your kernel could ever be.
→
The kernel engineer's skill mapCAREER
What hiring managers actually ask for — matmul-from-scratch to CUTLASS to AI-generated kernels — and where this site teaches each.
→
How to use this site
The two surfaces, the worklog style, the async track, and the order to read things in.
