papers/zoo-gpu-evm/sections/06-integration-with-zoo-l2.tex

\section{Integration with Zoo L2}

\label{sec:integration}

\subsection{VM Architecture}

Zoo L2 runs three specialized VMs alongside \cevm{}:

\begin{enumerate}[leftmargin=*]
\item \textbf{DEX VM}: Order matching, AMM swaps, liquidity provision. Highest conflict rates due to shared pool state.
\item \textbf{FHE VM}: Homomorphic encryption on ciphertext. Near-zero conflicts (isolated encrypted state).
\item \textbf{AI Mining VM}: Proof-of-useful-computation. Independent result slots per miner.
\end{enumerate}

Specialized VMs are implemented as EVM precompiles (\texttt{0x0300}--\texttt{0x0320}) dispatched to dedicated GPU kernels.

\subsection{Validator Requirements}

\begin{table}[h]
\centering
\begin{tabular}{lcc}
\toprule
\textbf{Tier} & \textbf{Minimum GPU} & \textbf{Stake} \\
\midrule
Full Validator & NVIDIA A100 (40GB) & 100M $\KEEPER$ \\
Light Validator & NVIDIA L4 (24GB) & 50M $\KEEPER$ \\
Archive Node & NVIDIA A100 (80GB) & --- \\
\bottomrule
\end{tabular}
\caption{Validator GPU and stake requirements}
\label{tab:validator-gpu}
\end{table}

Light validators execute only $C$ on GPU and verify $I$ via state-root comparison.

\subsection{Precompile Dispatch}

Precompile transactions are always placed in the conflict set to ensure execution against fully resolved base state:

\begin{algorithm}
\caption{Precompile-Aware Partitioning}
\label{alg:precompile}
\begin{algorithmic}[1]
\REQUIRE Transaction set $T$, precompile address set $P$
\STATE $T_{\text{pre}} \gets \{t \in T : \text{target}(t) \in P\}$
\STATE $I, C_{\text{evm}} \gets \text{Partition}(T \setminus T_{\text{pre}})$
\STATE $C \gets C_{\text{evm}} \cup T_{\text{pre}}$
\RETURN $I, C$
\end{algorithmic}
\end{algorithm}

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