papers/evmgpu-benchmark/sections/99-conclusion.tex

\section{Conclusion}


Measured benchmarks demonstrate four independent acceleration vectors, from Apple M1 Max Metal
to production NVIDIA GB10 and AMD RDNA silicon:

\begin{enumerate}
\item \textbf{GPU EVM kernel}: 85.9\,M opcodes/sec, 4.85$\times$ peak speedup over CPU.
      GPU wins above 500 opcodes/tx, reaching 3.09$\times$ at 10{,}000 ops/tx.
\item \textbf{C++ interpreter}: 20.9\,Ggas/s with full state (5.1$\times$ over Go EVM).
      416\,Ggas/s interpreter-only.
\item \textbf{Parallel ecrecover}: 7.1$\times$ on 10 cores. This is the highest-leverage
      single optimization because ecrecover consumes 87\% of block processing time.
\item \textbf{Production DEX matching}: on shipping accelerators, the \emph{deterministic}
      consensus-path matcher (one thread per book, coalesced and tuned per backend, GPU$=$CPU
      parity-verified) reaches \textbf{12.76\,billion orders/sec} on an AMD Radeon 8060S
      ($12.5\times$ its 32-core CPU) and \textbf{9.13\,billion orders/sec} on an NVIDIA GB10
      ($17.5\times$ its Grace CPU)---$5.9$--$6.6\times$ over each part's un-coalesced baseline---
      aggregating to \SI{21.9}{\giga\text{ord}/\second} across a two-node network.
\end{enumerate}

GPU memory scales linearly at 73\,KB/tx, supporting up to 800k transactions per block
on 64\,GB unified memory. The system is validated by 1{,}183 tests across C++, Rust, Go,
Metal, and WGSL.

These optimizations are orthogonal and composable: parallel ecrecover + C++ EVM +
GPU crypto + Block-STM can be combined for multiplicative throughput gains.

\paragraph{The unified engine, end to end.} Composing the measured layers---billion-order GPU
matching, a tuned \SI{2.35}{\giga bit/\second} inter-node link at 94\% of line rate, and
post-quantum ML-DSA-65 at \SI{134}{\micro\second} verification---a single colocated two-GPU
venue over \SI{800}{\giga bit/\second} InfiniBand admits an estimated
\SI{1.79}{\giga\text{orders}/\second}, exceeding the \emph{entire world's traded order flow} by
more than two orders of magnitude while remaining fully post-quantum and, modulo an estimated
$\sim$5--15\% confidential-compute overhead, fully confidential. Because the DEX matcher, the
native GPU EVM (\texttt{block\_stm\_execute\_kernel} and companions), and the 28-kernel FHE layer
(NTT primitive at $23.6\times$ over CPU) share one unified-memory address space and a single
dispatch stream, a confidential order is matched, settled, and homomorphically processed without
ever leaving the accelerator. This is the defining property of the Lux unified GPU-native
multichain engine: the empirical case that a post-quantum, confidential, GPU-native settlement
fabric can clear all of the world's markets from a single colocated site.

\begin{thebibliography}{9}

\bibitem{blockstm}
A.~Gelashvili, A.~Spiegelman, Z.~Xiang, G.~Danezis, Z.~Li, Y.~Malkhi, Y.~Xia, R.~Zhou.
\emph{Block-STM: Scaling Blockchain Execution by Turning Ordering Curse to a Performance Blessing}.
PPoPP 2023.

\bibitem{pevm}
Risechain.
\emph{pevm: Parallel EVM implementation in Rust}.
\url{https://github.com/risechain/pevm}, 2024.

\bibitem{reth}
Paradigm.
\emph{reth: Modular, contributor-friendly and blazing-fast implementation of the Ethereum protocol, in Rust}.
\url{https://github.com/paradigmxyz/reth}, 2023.

\bibitem{evmone}
ipsilon.
\emph{evmone: Fast Ethereum Virtual Machine implementation}.
\url{https://github.com/ipsilon/evmone}, 2019.

\bibitem{zap}
Hanzo AI.
\emph{ZAP: Zero-Copy Application Protocol}.
\url{https://github.com/zap-proto/zap}, 2025.

\bibitem{quasar}
Lux Industries, Inc.
\emph{Quasar: Post-Quantum Finality Consensus for Multi-Chain Networks}.
Lux Technical Report, 2025.

% --- Named-competitor scaffold (Apr 2026) ---

\bibitem{gethgithub}
Ethereum Foundation.
\emph{go-ethereum: Official Go implementation of the Ethereum protocol}.
\url{https://github.com/ethereum/go-ethereum}, 2014.

\bibitem{rethgithub}
Paradigm.
\emph{reth: Modular, contributor-friendly and blazing-fast implementation of the Ethereum protocol, in Rust}.
\url{https://github.com/paradigmxyz/reth}, 2023.

\bibitem{paradigmreth2024}
Bortolomeazzi, S. and Konstantopoulos, G.
\emph{Reth's path to 1 gigagas per second, and beyond}.
Paradigm, April 2024.
\url{https://www.paradigm.xyz/2024/04/reth-perf}

\bibitem{erigongithub}
Erigon Tech.
\emph{erigon: Ethereum execution client (Go)}.
\url{https://github.com/erigontech/erigon}

\bibitem{chainstackgetherigon}
Chainstack.
\emph{Erigon vs Geth: Which Ethereum client is better}.
\url{https://chainstack.com/ethereum-clients-geth-and-erigon/}, 2026.

\bibitem{evmonegithub}
ipsilon.
\emph{evmone: Fast Ethereum Virtual Machine implementation in C++}.
\url{https://github.com/ipsilon/evmone}, 2019.

\bibitem{blockstmppopp}
Gelashvili, R., Spiegelman, A., Xiang, Z., Danezis, G., Li, Z., Malkhi, D., Xia, Y., Zhou, R.
\emph{Block-STM: Scaling Blockchain Execution by Turning Ordering Curse to a Performance Blessing}.
PPoPP 2023.
arXiv:2203.06871.

\bibitem{pevmrise}
Risechain.
\emph{pevm: Parallel EVM implementation in Rust}.
\url{https://github.com/risechain/pevm}, 2024.

\bibitem{secp256k1lib}
Bitcoin Core.
\emph{libsecp256k1: Optimized C library for ECDSA signatures on the secp256k1 curve}.
\url{https://github.com/bitcoin-core/secp256k1}

\bibitem{supranationalblstevm}
Supranational.
\emph{blst: Multilingual BLS12-381 signature library}.
\url{https://github.com/supranational/blst}

\bibitem{knuth-taocp2}
Knuth, D.~E.
\emph{The Art of Computer Programming, Volume 2: Seminumerical Algorithms}.
Algorithm D (multiple-precision division), \S 4.3.1. Addison-Wesley.

\bibitem{lux-lightspeed-dex}
Lux Industries, Inc.
\emph{Lux Lightspeed DEX: On-Chain Order Book with Sub-Second Finality}.
Lux Technical Report.

\end{thebibliography}