papers/hanzo-fhe-inference/sections/08-production-results.tex

\section{Production Results}

\label{sec:results}
% ============================================================================

We benchmark three production workloads on H100 GPUs.

\begin{table}[H]
\centering
\small
\begin{tabular}{lrrrl}
\toprule
\textbf{Workload} & \textbf{Scheme} & \textbf{Latency} & \textbf{Accuracy} & \textbf{vs.\ Plaintext} \\
\midrule
Medical image (4-layer CNN) & CKKS & 1.8s & 97.1\% & $-$0.3\% \\
Credit scoring (XGBoost 500) & TFHE & 0.9s & 99.8\% & $-$0.0\% \\
Document embedding (2-layer) & CKKS & 0.6s & --- & cosine $>$ 0.998 \\
Fraud detection (RF 200) & TFHE & 0.4s & 99.9\% & $-$0.0\% \\
\bottomrule
\end{tabular}
\caption{Encrypted inference benchmarks on NVIDIA H100 GPU. Accuracy compared
to plaintext model evaluation.}
\label{tab:benchmarks}
\end{table}

CKKS-based neural network inference introduces minimal accuracy loss ($<$0.3\%)
from polynomial activation approximation. TFHE-based tree models preserve exact
accuracy since TFHE performs exact integer arithmetic.

\subsection{Throughput}

With GPU acceleration, a single H100 sustains:
\begin{itemize}[leftmargin=1.1em]
    \item CKKS: 35 encrypted neural network inferences per minute (4-layer,
    $N = 2^{15}$).
    \item TFHE: 67 encrypted XGBoost evaluations per minute (500 trees,
    depth 8, 16-bit features).
\end{itemize}

% ============================================================================