papers/zoo-gpu-evm/sections/07-benchmarks.tex

\section{Benchmarks}

\label{sec:benchmarks}

All benchmarks use identical transaction traces from Zoo L2 testnet (30 days, 42M transactions).

\subsection{Throughput}

\begin{table}[h]
\centering
\begin{tabular}{lcc}
\toprule
\textbf{Implementation} & \textbf{Throughput (Mops/sec)} & \textbf{Speedup} \\
\midrule
geth (Go, single-thread) & 97 & 1.0$\times$ \\
reth (Rust, single-thread) & 138 & 1.4$\times$ \\
evmone (C++, single-thread) & 162 & 1.7$\times$ \\
\cevm{} (C++/CUDA, H100) & 2,310 & 23.8$\times$ \\
\cevm{} (C++/CUDA, 4$\times$H100) & 8,890 & 91.6$\times$ \\
\bottomrule
\end{tabular}
\caption{Throughput comparison across EVM implementations}
\label{tab:throughput}
\end{table}

\subsection{Speedup by Transaction Type}

\begin{table}[h]
\centering
\begin{tabular}{lccc}
\toprule
\textbf{Type} & \textbf{CPU (Mops/s)} & \textbf{\cevm{} (Mops/s)} & \textbf{Speedup} \\
\midrule
EOA transfers & 310 & 4,820 & 15.5$\times$ \\
ERC-20 transfers & 145 & 3,410 & 23.5$\times$ \\
DEX swaps & 82 & 1,240 & 15.1$\times$ \\
AI Mining submissions & 61 & 2,890 & 47.4$\times$ \\
FHE operations & 23 & 1,070 & 46.5$\times$ \\
Complex contract chains & 44 & 680 & 15.5$\times$ \\
\bottomrule
\end{tabular}
\caption{Per-type speedup (single H100)}
\label{tab:type-speedup}
\end{table}

AI Mining and FHE achieve $>$45$\times$ speedups due to near-zero conflict rates.
DEX swaps achieve 15$\times$ due to shared pool state.

\subsection{Latency}

\begin{table}[h]
\centering
\begin{tabular}{lcc}
\toprule
\textbf{Phase} & \textbf{CPU (ms)} & \textbf{\cevm{} H100 (ms)} \\
\midrule
Dependency analysis & --- & 1.2 \\
State snapshot to GPU & --- & 3.8 \\
Parallel execution & --- & 18.4 \\
Write-log merge + re-exec & --- & 5.0 \\
Sequential execution (all) & 847.0 & --- \\
State root computation & 12.3 & 12.3 \\
\midrule
\textbf{Total (10K txns)} & \textbf{859.3} & \textbf{40.7} \\
\bottomrule
\end{tabular}
\caption{Per-block latency breakdown}
\label{tab:latency}
\end{table}

All \cevm{} latencies remain within the 12-second \Lux{} slot time at all tested block sizes (1K--20K transactions).

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