1\begin{table}[t]2\centering\small3\caption{Gate F: Qwen3-30B-A3B on M3U96a (bounded residency, cold NVMe) with M3U96b as expert worker over 1 GbE; placements: none (all local), odd (half remote, static), popular (least popular half remote, frequency-based); 64 tokens, held-out general/code, 8 runs per row.}4\label{tab:gateF}5\begin{tabular}{llrrrrrr}6\toprule7Backend & cfg & $n$ & decode tok/s (median) & p95 ITL ms & hit & stall s & TTFT ms \\8\midrule9titan-distributed & C=32, placement=none, cold & 16 & 8.8 & 187.7 & 0.775 & 4.23 & 831 \\10titan-distributed & C=32, placement=odd, cold & 16 & 4.2 & 263.7 & 0.922 & 1.36 & 1054 \\11titan-distributed & C=32, placement=popular, cold & 16 & 5.1 & 236.2 & 0.877 & 2.18 & 731 \\12titan-distributed & C=64, placement=none, cold & 16 & 16.3 & 100.2 & 0.922 & 1.37 & 470 \\13titan-distributed & C=64, placement=odd, cold & 16 & 5.2 & 224.1 & 1.000 & 0.00 & 927 \\14titan-distributed & C=64, placement=popular, cold & 16 & 6.0 & 200.9 & 1.000 & 0.00 & 714 \\15\bottomrule16\end{tabular}17\begin{tablenotes}\footnotesize18\item Sources: \code{benchmarks/raw/gateF/two\_node\_qwen3\_m3u96a\_m3u96b.jsonl.tmp}.19\end{tablenotes}20\end{table}21