papers/gpu-evm-whitepaper/sections/09-comparison-with-external-gpu-evm-project.tex

\section{Comparison with External GPU EVM Projects}

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

Several projects have explored GPU-accelerated EVM execution. We compare cevm against
the most relevant:

\subsection{CuEVM (FaceBlockTeam/gpuEVM)}

CuEVM~\cite{cuevm} is a CUDA-based GPU EVM implementation using CGBN (Cooperative Groups
Big Number) for 256-bit arithmetic. Key characteristics:

\begin{itemize}
\item \textbf{Backend}: CUDA only (NVIDIA GPUs). Uses cooperative groups for 256-bit
      integer operations, requiring warp-level synchronization.
\item \textbf{Performance}: 60--100K TPS experimental, primarily benchmarked on simple
      transfer workloads.
\item \textbf{Target}: Designed for EVM fuzzing and testing, not production block execution.
      Lacks conflict detection (no Block-STM or equivalent).
\item \textbf{State}: No GPU-resident state---state is transferred from CPU per block.
\end{itemize}

\subsection{GatlingX}

GatlingX~\cite{gatlingx} claims 200$\times$ speedup for GPU-accelerated EVM but provides
no reproducible benchmarks, no open source implementation, and no peer-reviewed evaluation.
We do not consider unverifiable claims.

\subsection{cevm Differentiation}

\begin{table}[h]
\centering
\small
\begin{tabular}{lllll}
\toprule
\textbf{Feature} & \textbf{cevm} & \textbf{CuEVM} & \textbf{GatlingX} \\
\midrule
GPU backend         & Metal + CUDA + Dawn/WebGPU & CUDA only     & Unknown \\
Block-STM on GPU    & Yes (novel)             & No               & Unknown \\
GPU-resident state  & Yes (hash table)        & No               & Unknown \\
Production target   & Yes                     & No (fuzzing)     & Claimed \\
Conflict detection  & GPU atomics             & None             & Unknown \\
Open benchmarks     & This paper              & Partial          & None \\
PQ crypto on GPU    & Yes (ML-DSA/ML-KEM/SLH) & No               & No \\
FHE on GPU          & Yes (TFHE)              & No               & No \\
\bottomrule
\end{tabular}
\caption{Feature comparison of GPU EVM implementations. cevm ships three backends
(Metal, CUDA, Dawn/WebGPU) with CPU SIMD fallback, and is the only implementation
with GPU-resident Block-STM, persistent GPU state, and post-quantum/FHE acceleration.}
\label{tab:gpu_comparison}
\end{table}

The key novelty of cevm is \textbf{Metal Block-STM on GPU}: no published work combines
the Block-STM optimistic concurrency protocol~\cite{blockstm} with GPU execution. CuEVM
executes transactions independently without conflict detection, which is sufficient for
fuzzing but not for production block processing where transactions share state.

\subsection{Comparison with CPU-Side Parallel EVM}

\begin{table}[h]
\centering
\small
\begin{tabular}{lrrp{4cm}}
\toprule
\textbf{System} & \textbf{Throughput} & \textbf{Type} & \textbf{Notes} \\
\midrule
Ethereum mainnet    & $\sim$30\,Mgas/s     & Production & Sequential geth, 12s blocks, $\sim$30M gas/block \\
Lux C-Chain (Go)    & 534--1{,}038\,Mgas/s & Measured   & Sequential, execution only \\
Solana              & $\sim$50k TPS        & Production & Not EVM; SVM/BPF runtime \\
Monad (claimed)     & 10{,}000 TPS         & Claimed    & Proprietary parallel EVM, not open source \\
pevm (Rust)         & 19--22\,Ggas/s       & Published  & Parallel Block-STM, ETH transfers \\
CuEVM (CUDA)        & 60--100K TPS         & Experimental & Fuzzing target, no conflict detection \\
\midrule
\textbf{cevm} C++   & 20.9\,Ggas/s        & Measured   & EVM-only, full state \\
\textbf{cevm} GPU   & 85.9\,M ops/s       & Measured   & GPU kernel, 4.85$\times$ \\
\textbf{cevm} Block-STM & 3{,}390\,Mgas/s & Measured   & 10-thread, mixed workload \\
\bottomrule
\end{tabular}
\caption{Throughput comparison with external systems. Ethereum and Lux numbers are
directly comparable (both EVM). Solana uses a different VM (SVM) and is not directly
comparable. Monad's numbers are self-reported and not independently verified.
pevm numbers from published benchmarks on ETH transfers (zero-conflict, best case).
CuEVM numbers from~\cite{cuevm}.}
\label{tab:external}
\end{table}

\textbf{Key observations:}
\begin{itemize}
\item cevm C++ matches pevm Rust on EVM-only throughput ($\sim$20\,Ggas/s vs 19--22\,Ggas/s),
      confirming that the execution engine is not the differentiator at this performance level.
\item cevm's GPU acceleration (4.85$\times$) applies \emph{on top of} the C++ interpreter,
      targeting the opcode dispatch that C++ and Rust handle equally well.
\item The real differentiator is the GPU pipeline: crypto offloading (ecrecover, Keccak-256,
      PQ signatures) addresses the 87\% of time that pure interpreter optimization cannot touch.
\item CuEVM's lack of conflict detection limits it to embarrassingly parallel workloads
      (fuzzing, independent transfers). cevm's GPU Block-STM handles real-world DeFi workloads
      with shared state.
\end{itemize}

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