Ruach Tov Collective — Research Dashboards
🎯 KernelBench Coverage (L1 / L2 / L3)
Stanford KernelBench verification: single kernels (L1/L2) and multi-kernel compositions (L3). Every cell shows its instrument and auto-timestamp on hover.
KernelBench Coverage Grid
Table 10000
L1 (100/100 auto-lift), L2 (100/100 whole-model bit-exact), and L3 (20/23 store-units bit-exact) on one grid. Click cells for source drill-down. Hover for auto-timestamps.
Improvement Scoreboard
28 independently verified 0-ULP improvements (17 CPU + 11 CUDA). Speedup ranges, not single numbers. Sorted by guaranteed floor.
🔬 BPD Kernel Fusion Programme
How Does Our GPU Actually Work?
Measured probe kernel results on Tesla P4 (sm_61): memory floor decomposition, three compute regimes, in-place vs two-buffer (2.6× at L2 boundary), cache hierarchy sweeps, launch overhead, coalescing impact. Every number measured from Prolog via CUPTI hardware counters.
Detailed dashboards from the kernel fusion verification campaign.
LLVM IR Emitter Match Status
Table 10001
ULP match between LLVM IR emitters and ggml reference. 6 ops at 0-ULP (blue), per-op ULP distances.
📐 Sorting Networks
Papers & Catalogs
Published research: Broken Lattice book (88pp), Précis Volume I, n=32 Untangled, Mutex Comparator, and more.
