1\begin{table}[t]2\centering\small3\caption{H4: fused up+gate gather\_qmm vs two kernels (laptop; 64 slots resident, 128 tokens, 12 runs; 16 slots cold, 64 tokens, 4 runs).}4\label{tab:fused}5\begin{tabular}{llrrrrrr}6\toprule7Backend & cfg & $n$ & decode tok/s (median) & p95 ITL ms & hit & stall s & TTFT ms \\8\midrule9titan & C=16, fused=False, cold & 4 & 35.0 & 36.9 & 0.492 & 1.17 & 348 \\10titan & C=16, fused=True, cold & 4 & 35.9 & 35.0 & 0.492 & 1.10 & 342 \\11titan & C=64, fused=False & 12 & 476.7 & 2.2 & 1.000 & 0.00 & 17 \\12titan & C=64, fused=True & 12 & 475.6 & 2.2 & 1.000 & 0.00 & 18 \\13\bottomrule14\end{tabular}15\begin{tablenotes}\footnotesize16\item Sources: \code{benchmarks/raw/gateD/fused\_upgate\_laptop.jsonl}.17\end{tablenotes}18\end{table}19