papers/lux-performance-security-tradeoffs/sections/10-benchmark-summary.tex

\section{Benchmark Summary}


Complete benchmark table, all measured on production or equivalent hardware:

\begin{center}
\begin{tabular}{lrrrr}
\toprule
\textbf{Operation} & \textbf{Time} & \textbf{Memory} & \textbf{Output Size} & \textbf{PQ} \\
\midrule
\multicolumn{5}{l}{\textit{Signatures}} \\
ECDSA sign & 12$\mu$s & 72 B & 64 B & No \\
ECDSA verify & 45$\mu$s & 72 B & --- & No \\
BLS sign & 15$\mu$s & 96 B & 96 B & No \\
BLS aggregate ($N$=100) & 10$\mu$s & 48 B & 48 B & No \\
BLS verify aggregate & 1.2ms & 96 B & --- & No \\
ML-DSA-65 sign & 420$\mu$s & 44 KB & 3,309 B & Yes \\
ML-DSA-65 verify & 180$\mu$s & 22 KB & --- & Yes \\
ML-DSA-65 keygen & 678$\mu$s & 88 KB & 1,952 B (pk) & Yes \\
\midrule
\multicolumn{5}{l}{\textit{Aggregation}} \\
STARK gen ($N$=100) & 580ms & 350 MB & $\sim$200 B & Yes$^*$ \\
STARK verify ($N$=100) & 3.2ms & 2 MB & --- & Yes$^*$ \\
\midrule
\multicolumn{5}{l}{\textit{Threshold}} \\
FROST DKG (5-of-9) & 2.1s & 12 MB & Key shares & No$^{**}$ \\
FROST sign (3-of-5) & 85ms & 1 MB & 64 B & No$^{**}$ \\
FROST sign (5-of-9) & 178ms & 2 MB & 64 B & No$^{**}$ \\
\midrule
\multicolumn{5}{l}{\textit{FHE}} \\
TFHE bootstrap & 12ms & 24 MB & 4 KB & Yes \\
TFHE 16-bit compare & 380ms & 48 MB & 4 KB & Yes \\
TFHE order match & 790ms & 96 MB & 8 KB & Yes \\
\midrule
\multicolumn{5}{l}{\textit{Consensus}} \\
QuasarCert verify (CPU) & ${\sim}2$--$5$\,ms & --- & 248 B & Yes \\
\bottomrule
\end{tabular}
\end{center}

$^*$STARK itself is PQ (hash-based); the ML-DSA statements inside are PQ.\\
$^{**}$FROST over Schnorr/secp256k1 is not PQ; FROST can be instantiated over lattice-based schemes for PQ (Pulsar).