Kernel Climb
Every Open Kernel Challenge kernel board drawn as hill climbing over its real dimensions. Every point is a passed scored run and every height is measured. The ceilings are real: the H200's hardware limit in both modes, and in "With the field" mode, where the leaders finished when the kernel track closed on Nov 4.
The climb
The hardware limit, in one example. On the decode layer's batch-1 sliding-window shape, SGLang takes about 173 µs. Just reading that step's weights and KV cache once, at the H200's full 4.8 TB/s, takes 40 µs. So no kernel can beat SGLang there by more than about 4.3×. The judge publishes this floor for every shape: the time to move the data at 4.8 TB/s or do the math at 989 TFLOP/s, whichever is longer. Putting every shape on its floor gives each board's limit.
Per-shape terrain
One bar per scored shape, for the run on the scrubber. Loads run left to right and the shapes inside each load front to back. The grid at 1.0× is SGLang. The small diamonds mark the board's first scored run. Switch the height to see how close each shape gets to its hardware limit (100% would be the H200 at full memory speed).
Run ledger
| date (UTC) | run | step | score | Δ | A | B | note |
|---|