papers/zoo-ai-memory/sections/10-benchmarks.tex

\section{Benchmarks}

\label{sec:benchmarks}
%----------------------------------------------------------------------

All benchmarks measured on \Zoo{} Network testnet with 21 validators, each running on AMD EPYC 7763 with 256 GB RAM and NVIDIA A100 80 GB.

\subsection{Core Operations}

\begin{table}[h]
\centering
\begin{tabular}{lcc}
\toprule
\textbf{Operation} & \textbf{Latency (p50)} & \textbf{Latency (p99)} \\
\midrule
Entry append (1 KB) & 1.2 ms & 3.8 ms \\
Entry append (32 KB) & 4.7 ms & 12.1 ms \\
Vault read (single entry) & 0.3 ms & 1.1 ms \\
Similarity search (10k entries) & 8.4 ms & 22.7 ms \\
CRDT merge (100 ops) & 2.1 ms & 6.3 ms \\
CRDT merge (10k ops) & 148 ms & 312 ms \\
\IChain{} anchor update & 1.8 s & 3.2 s \\
\bottomrule
\end{tabular}
\caption{Core operation latencies}
\label{tab:core_bench}
\end{table}

\subsection{CRDT Synchronization}

\begin{table}[h]
\centering
\begin{tabular}{lcc}
\toprule
\textbf{Divergence (ops)} & \textbf{Sync Time} & \textbf{Throughput} \\
\midrule
100 & 7.1 ms & 14,084 ops/s \\
1,000 & 68 ms & 14,705 ops/s \\
10,000 & 712 ms & 14,044 ops/s \\
100,000 & 7.4 s & 13,513 ops/s \\
\bottomrule
\end{tabular}
\caption{CRDT sync throughput}
\label{tab:crdt_bench}
\end{table}

\subsection{FHE Operations}

\begin{table}[h]
\centering
\begin{tabular}{lccc}
\toprule
\textbf{Operation} & \textbf{CPU} & \textbf{GPU (A100)} & \textbf{Speedup} \\
\midrule
CKKS encrypt (768-dim) & 12.4 ms & 0.31 ms & 40$\times$ \\
CKKS inner product & 28.6 ms & 0.72 ms & 39.7$\times$ \\
Encrypted recall (1k entries) & 34.2 s & 0.86 s & 39.8$\times$ \\
Encrypted recall (10k entries) & 341 s & 8.6 s & 39.7$\times$ \\
TFHE comparison (single) & 8.3 ms & 0.21 ms & 39.5$\times$ \\
Threshold reveal (67-of-100) & 240 ms & 240 ms & 1$\times$ \\
\bottomrule
\end{tabular}
\caption{FHE operation benchmarks (GPU acceleration via \Zoo{} \AChain{})}
\label{tab:fhe_bench}
\end{table}

The threshold reveal latency is dominated by network round-trips to \TChain{} validators, not computation, hence no GPU speedup.

\subsection{End-to-End Memory Recall}

Complete encrypted memory recall (query $\to$ FHE search $\to$ threshold decrypt $\to$ return) for a 10k-entry capsule:

\begin{itemize}
    \item \textbf{Unencrypted baseline:} 8.4 ms (similarity search only)
    \item \textbf{FHE + GPU + threshold:} 8.6 s + 240 ms = 8.84 s
    \item \textbf{Overhead factor:} $\sim$1,050$\times$
\end{itemize}

This overhead is acceptable for high-sensitivity applications where privacy justifies latency. For most use cases, standard encrypted vault access (8.4 ms) suffices.

%----------------------------------------------------------------------