papers/zoo-quasar-benchmarks/sections/05-poai-overhead-analysis.tex

\section{PoAI Overhead Analysis}

\label{sec:poai-overhead}

\subsection{Per-Block Overhead}

\begin{table}[t]
\centering
\begin{tabular}{lccc}
\toprule
\textbf{Component} & \textbf{CPU (ms)} & \textbf{GPU (ms)} & \textbf{Total (ms)} \\
\midrule
Attestation parsing & 0.05 & --- & 0.05 \\
TEE signature check & 0.25 & --- & 0.25 \\
BLS aggregation & 0.10 & --- & 0.10 \\
\midrule
\textbf{Per-block total} & \textbf{0.40} & --- & \textbf{0.40} \\
\bottomrule
\end{tabular}
\caption{PoAI per-block overhead. Only attestation verification runs per block.}
\label{tab:per-block}
\end{table}

\subsection{Per-Epoch Overhead}

\begin{table}[t]
\centering
\begin{tabular}{lccc}
\toprule
\textbf{Component} & \textbf{CPU (ms)} & \textbf{GPU (ms)} & \textbf{Total (ms)} \\
\midrule
Quality evaluation & 15 & 8 & 23 \\
Reward computation & 3 & --- & 3 \\
Weight update & 2 & --- & 2 \\
Epoch finalization & 5 & --- & 5 \\
\midrule
\textbf{Per-epoch total} & \textbf{25} & \textbf{8} & \textbf{33} \\
\bottomrule
\end{tabular}
\caption{PoAI per-epoch overhead. Epoch = 100 blocks $\approx$ 630ms at 21 validators.}
\label{tab:per-epoch}
\end{table}

The per-epoch overhead of 33ms is amortized across 100 blocks, contributing $33/100 = 0.33$ms per block. Combined with the per-block overhead of 0.40ms, the total \PoAI{} overhead is \textbf{0.73ms per block} or approximately \textbf{7.3\% of \Quasar{} finality latency}.

\subsection{Impact on Useful Compute}

\begin{table}[t]
\centering
\begin{tabular}{lccc}
\toprule
\textbf{Metric} & \textbf{Quasar Only} & \textbf{Quasar + PoAI} & \textbf{Overhead} \\
\midrule
Finality (p50) & 6.3ms & 6.7ms & +6.3\% \\
Finality (p99) & 12.4ms & 13.2ms & +6.5\% \\
TPS (transfers) & 4,760,000 & 4,712,000 & $-$1.0\% \\
TPS (mixed) & 3,210,000 & 3,130,000 & $-$2.5\% \\
Useful compute & 0\% & 98.2\% & --- \\
\bottomrule
\end{tabular}
\caption{Impact of \PoAI{} on \Quasar{} performance. Overhead is minimal; useful compute gain is substantial.}
\label{tab:useful}
\end{table}

The dual-layer design achieves 98.2\% useful compute utilization: 98.2\% of validator GPU cycles are spent on AI inference, training, or data processing, with only 1.8\% consumed by consensus and verification overhead.