papers/zoo-evm-benchmarks/sections/04-poai-consensus-overhead.tex

\section{PoAI Consensus Overhead}

\label{sec:poai}

Zoo Network runs Proof of AI~\cite{zoo-poai-consensus} as a consensus overlay on Quasar~\cite{lux-quasar-consensus}. PoAI adds verification of AI compute contributions, introducing overhead beyond pure Quasar finality.

\subsection{Overhead Measurement}

\begin{table}[t]
\centering
\begin{tabular}{lccc}
\toprule
\textbf{Component} & \textbf{Latency (ms)} & \textbf{Overhead} \\
\midrule
Quasar consensus (baseline) & 10.0 & --- \\
TEE attestation verification & 0.3 & 3.0\% \\
Quality scoring (inference) & 0.5 & 5.0\% \\
BLS aggregation & 0.1 & 1.0\% \\
\midrule
\textbf{Total with PoAI} & \textbf{10.9} & \textbf{9.0\%} \\
\bottomrule
\end{tabular}
\caption{PoAI consensus overhead breakdown. TEE attestation is the dominant cost.}
\label{tab:poai}
\end{table}

The 9\% overhead is a worst case when every block contains AI compute attestations. For blocks with only financial transactions (no AI workload), the PoAI overhead drops to 1\% (BLS aggregation only).

\subsection{Throughput Under PoAI}

\begin{table}[t]
\centering
\begin{tabular}{lcc}
\toprule
\textbf{Workload Mix} & \textbf{TPS (Quasar only)} & \textbf{TPS (Quasar + PoAI)} \\
\midrule
100\% transfers & 4,760,000 & 4,712,000 \\
70\% transfers / 30\% AI & 3,200,000 & 2,944,000 \\
50\% transfers / 50\% AI & 2,380,000 & 2,190,000 \\
100\% AI inference & 200,000 & 184,000 \\
\bottomrule
\end{tabular}
\caption{Throughput impact of PoAI verification for mixed workloads.}
\label{tab:poai-throughput}
\end{table}