{ "provenance": "Hand-written example kernels in examples/ scored with `cakebench suite --device cuda --full` on one shared A100-SXM4-40GB, as recorded in this repo's README at commit 7631683. Timing is median GPU span with L2 flushed; shared box, so timings are noisy.", "rows": [ {"task": "kmeans_assign", "shapes_correct": 3, "shapes_total": 3, "geomean_speedup": 1.69}, {"task": "knn_search", "shapes_correct": 3, "shapes_total": 3, "geomean_speedup": 1.13}, {"task": "tinygemm", "shapes_correct": 8, "shapes_total": 8, "geomean_speedup": 1.00}, {"task": "kda_decode", "shapes_correct": 4, "shapes_total": 4, "geomean_speedup": 1.03}, {"task": "kda_prefill", "shapes_correct": 3, "shapes_total": 3, "geomean_speedup": 1.06} ] }