papers/evmgpu-benchmark/sections/10-discussion.tex

\section{Discussion}


\subsection{When to Use GPU EVM}

The crossover analysis (Table~\ref{tab:crossover}) provides a clear decision rule:

\begin{itemize}
\item \textbf{Below 500 ops/tx}: CPU wins. Simple transfers and token operations
      do not generate enough parallel work to amortize GPU launch overhead.
\item \textbf{500--2{,}000 ops/tx}: Marginal GPU advantage (1.0--2.08$\times$).
      Worth offloading only if the GPU is not otherwise occupied.
\item \textbf{Above 2{,}000 ops/tx}: Clear GPU advantage. Uniswap swaps ($\sim$3{,}000 ops),
      complex DeFi ($>$5{,}000 ops), and ZK verification ($>$10{,}000 ops) all benefit.
\item \textbf{Peak}: 50k transactions $\times$ 5k opcodes each: 4.85$\times$.
\end{itemize}

\subsection{The Optimization Priority Stack}

Based on measured data, the priority order for EVM acceleration is:

\begin{enumerate}
\item \textbf{Parallel ecrecover} (7.1$\times$ on 87\% of time = 3.76$\times$ overall).
      Implemented. Deployed.
\item \textbf{C++ interpreter} (5.1$\times$ on 13\% of time = 1.12$\times$ overall).
      Implemented. The marginal gain is small because EVM execution is not the bottleneck.
\item \textbf{GPU opcodes} (4.85$\times$ on complex contracts). Highest impact on
      computation-heavy workloads where per-transaction opcode count exceeds 2{,}000.
\item \textbf{Block-STM} (6.4$\times$ on 10 threads for independent transactions).
      Orthogonal to the above; parallelizes across transactions rather than within them.
\end{enumerate}

\subsection{Memory Budget}

GPU memory is the practical scaling constraint. At 73\,KB per transaction plus 5\,GB
persistent state, the memory budget determines maximum block size:

\begin{itemize}
\item \textbf{32\,GB system}: $(32 - 5) / 0.073 \approx 370{,}000$ tx/block
\item \textbf{64\,GB system}: $(64 - 5) / 0.073 \approx 800{,}000$ tx/block
\item \textbf{128\,GB (M4 Ultra)}: $(128 - 5) / 0.073 \approx 1{,}685{,}000$ tx/block
\end{itemize}

\subsection{Comparison to External Systems}

\begin{table}[h]
\centering
\small
\begin{tabular}{lrrl}
\toprule
\textbf{System} & \textbf{Throughput} & \textbf{Notes} \\
\midrule
Ethereum mainnet         & $\sim$30\,Mgas/s   & Sequential geth \\
Go EVM (Lux, sequential) & 534--1{,}038\,Mgas/s & Measured (this paper) \\
C++ evmone++ (Lux)       & 20.9\,Ggas/s       & EVM-only, measured \\
Go EVM + parallel sig    & $\sim$2{,}150\,Mgas/s & Measured (this paper) \\
Block-STM 10-thread      & 3{,}390\,Mgas/s    & Measured (this paper) \\
GPU EVM (50k$\times$5k)  & 85.9\,M ops/s      & Measured (this paper) \\
\bottomrule
\end{tabular}
\caption{Throughput comparison. Ethereum mainnet number is the effective gas
rate under consensus ($\sim$30M gas per 12s block). Lux numbers are execution-only,
measured without consensus overhead.}
\label{tab:comparison}
\end{table}

\subsection{Go EVM Workload Benchmarks at 1\,B Gas Blocks}

With 47K transactions per block ($\sim$1\,B gas at 21{,}000 gas/tx), the Go EVM
produces detailed per-workload timing that separates EVM execution from signature
recovery:

\begin{table}[h]
\centering
\small
\begin{tabular}{lrrrr}
\toprule
\textbf{Workload} & \textbf{EVM (ms)} & \textbf{Sig (ms)} & \textbf{Total (ms)} & \textbf{EVM-only (Mgas/s)} \\
\midrule
ETH Transfer    & 241  & 1{,}613 & 1{,}854 & \num{4148} \\
ERC-20 Transfer & 326  & 1{,}614 & 1{,}940 & \num{4537} \\
Storage Write   & 384  & 1{,}623 & 2{,}007 & \num{5970} \\
Uniswap Swap    & 368  & 1{,}624 & 1{,}992 & \num{4669} \\
Mixed DeFi      & 312  & 1{,}622 & 1{,}934 & \num{4723} \\
\bottomrule
\end{tabular}
\caption{Go EVM workload benchmarks at 47K txs/block (1\,B gas). Signature recovery
(\SI{1613}{ms}) is constant across workloads---it dominates total time at 82--87\%.
EVM-only throughput ranges from \SI{4148}{Mgas/s} (ETH Transfer) to \SI{5970}{Mgas/s}
(Storage Write). All measurements on Apple M1 Max.}
\label{tab:workload}
\end{table}

The signature bottleneck is stark: \SI{1613}{ms} of ecrecover is constant regardless of
workload complexity, confirming that GPU batch ecrecover is the highest-leverage
optimization at production block sizes.

\subsection{Consensus Pipeline Benchmarks}

With pipelined consensus (building block $N{+}1$ while block $N$ finalizes), the
consensus layer is no longer the throughput bottleneck:

\begin{table}[h]
\centering
\small
\begin{tabular}{lrrr}
\toprule
\textbf{Configuration} & \textbf{Blocks/sec} & \textbf{Txs/block} & \textbf{Theoretical TPS} \\
\midrule
SoloGPU ($K{=}1$)  & \num{107000} & 47K  & \num{5100000000} \\
BurstParams         & \num{78000}  & 100K & \num{7800000000} \\
\bottomrule
\end{tabular}
\caption{Consensus pipeline throughput on Apple M1 Max. SoloGPU uses $K{=}1$
(single validator, no network RTT), 1\,ms block time, 1\,B gas limit.
BurstParams uses 100K txs/block with 2.1\,B gas limit. These are consensus-layer
measurements---execution is excluded. Theoretical TPS = blocks/sec $\times$ txs/block.}
\label{tab:consensus_pipeline}
\end{table}

\textbf{Key finding}: Consensus is no longer the bottleneck. At \num{107000} blocks/sec
(SoloGPU) and \num{78000} blocks/sec (BurstParams), the consensus layer vastly outpaces
EVM execution. The bottleneck has shifted entirely to execution and signature recovery.

\subsection{Throughput at 1\,B Gas Block Limit}

Lux Network operates with a \num{1000000000} (1\,B) gas block limit---33$\times$ Ethereum's
30\,M limit. At 21{,}000 gas per simple transfer, this yields $\sim$47{,}619 transactions
per block. The following table projects execution throughput at this block size.

\begin{table}[h]
\centering
\small
\begin{tabular}{lrrr}
\toprule
\textbf{Execution Engine} & \textbf{Time/1B-gas block} & \textbf{Exec TPS$^*$} & \textbf{Notes} \\
\midrule
C++ evmone++ (1 core)    & 47.8\,ms  & \num{952000}     & $1\text{B} / 20.9\,\text{Ggas/s}$ \\
GPU EVM (Metal, 32-core) & $\sim$10\,ms & $\sim$\num{47600000} & Projected from 2.3\,Gops/s \\
Block-STM (10 threads)   & 295\,ms   & \num{161000}     & $1\text{B} / 3.39\,\text{Ggas/s}$ \\
Go EVM (sequential)      & 1{,}852\,ms & \num{25700}    & $1\text{B} / 0.54\,\text{Ggas/s}$ \\
\bottomrule
\end{tabular}
\caption{Execution throughput projected to Lux's 1\,B gas block limit.
$^*$Execution-only TPS assuming 21{,}000 gas/tx and no consensus overhead.
C++ evmone++ processes a full 1\,B-gas block in under 50\,ms, enabling
sub-100\,ms block times. GPU EVM projection assumes linear scaling from
measured 2.3\,Gops/s kernel throughput.}
\label{tab:1b-gas}
\end{table}

With BLS consensus aggregate verify at $16.51\times$ at $n{=}1024$
on the same-message hot path (cevm v0.47.1, pubkey-decompression cache
on top of v0.45 batching) and pipelined block execution, the
theoretical throughput ceiling is:
\begin{equation}
\text{TPS}_{\text{ceiling}} = \frac{47{,}619\text{ txs/block}}{10\,\text{ms}} = \num{4761904}\text{ TPS (with finality)}
\end{equation}
With 1\,ms InfiniBand interconnect replacing WAN latency, the theoretical
maximum reaches $\sim$\num{47619047} TPS.