mirror of
https://github.com/hanzoai/papers.git
synced 2026-10-29 07:34:31 +00:00
91 lines
2.9 KiB
TeX
91 lines
2.9 KiB
TeX
\section{Evaluation}
|
|
|
|
\label{sec:evaluation}
|
|
|
|
\subsection{Testnet Configuration}
|
|
|
|
We deployed ACI on a testnet with:
|
|
|
|
\begin{itemize}[leftmargin=1.1em]
|
|
\item \textbf{Validators}: 100 nodes across 10 geographic regions.
|
|
\item \textbf{Workers}: 500 GPU nodes (mixed A100, H100, RTX 4090).
|
|
\item \textbf{Clients}: 1,000 simulated agents submitting Poisson-distributed workloads.
|
|
\item \textbf{Duration}: 30 days of continuous operation.
|
|
\end{itemize}
|
|
|
|
\subsection{Throughput and Latency}
|
|
|
|
\begin{table}[H]
|
|
\centering
|
|
\small
|
|
\begin{tabular}{lcc}
|
|
\toprule
|
|
\textbf{Metric} & \textbf{ACI} & \textbf{Ethereum L2} \\
|
|
\midrule
|
|
Job matches/second & 12,000 & 350 \\
|
|
Settlement TPS & 2,000 & 150 \\
|
|
Match latency (P50) & 87ms & 2,400ms \\
|
|
Match latency (P99) & 198ms & 8,700ms \\
|
|
Settlement finality & 2.0s & 12s \\
|
|
\bottomrule
|
|
\end{tabular}
|
|
\caption{Throughput and latency comparison.}
|
|
\label{tab:throughput}
|
|
\end{table}
|
|
|
|
\subsection{PoAI Verification Results}
|
|
|
|
\begin{table}[H]
|
|
\centering
|
|
\small
|
|
\begin{tabular}{lcc}
|
|
\toprule
|
|
\textbf{Metric} & \textbf{Value} \\
|
|
\midrule
|
|
Total jobs processed & 8,412,893 \\
|
|
Jobs sampled for verification & 25,238 (0.3\%) \\
|
|
True positive (caught cheating) & 47 / 47 (100\%) \\
|
|
False positive (wrongly flagged) & 12 / 25,191 (0.05\%) \\
|
|
Verification overhead & 0.31\% of total compute \\
|
|
Detection rate (injected faults) & 99.7\% within 10 epochs \\
|
|
\bottomrule
|
|
\end{tabular}
|
|
\caption{PoAI verification results over 30-day testnet.}
|
|
\label{tab:poai_results}
|
|
\end{table}
|
|
|
|
We injected 47 deliberately faulty jobs (incorrect outputs) across 10 workers. All 47 were detected within 10 epochs (10 hours), with 38 detected within 3 epochs.
|
|
|
|
\subsection{Gas Cost Comparison}
|
|
|
|
\begin{table}[H]
|
|
\centering
|
|
\small
|
|
\begin{tabular}{lccc}
|
|
\toprule
|
|
\textbf{Operation} & \textbf{ACI Gas} & \textbf{Eth L2 Gas} & \textbf{Savings} \\
|
|
\midrule
|
|
Job submission & 42K & 147K & 71\% \\
|
|
Settlement & 28K & 89K & 69\% \\
|
|
Dispute resolution & 120K & 412K & 71\% \\
|
|
HMM swap & 35K & 118K & 70\% \\
|
|
Stake/unstake & 21K & 67K & 69\% \\
|
|
\midrule
|
|
\textbf{Average} & \textbf{49K} & \textbf{167K} & \textbf{73\%} \\
|
|
\bottomrule
|
|
\end{tabular}
|
|
\caption{Gas cost comparison between ACI and Ethereum L2.}
|
|
\label{tab:gas}
|
|
\end{table}
|
|
|
|
The 73\% gas savings come from AIVM's native tensor operations and verification opcodes, which would require expensive general-purpose EVM operations on Ethereum.
|
|
|
|
\subsection{Stress Testing}
|
|
|
|
\begin{enumerate}[leftmargin=1.4em]
|
|
\item \textbf{10x load surge}: Network maintained 99.8\% job success rate under 10x normal load for 1 hour. Fast-path latency increased to 420ms (from 87ms), but remained within SLA.
|
|
\item \textbf{33\% validator failure}: With 33 of 100 validators offline, the network continued operating with degraded settlement throughput (1,200 TPS, down from 2,000).
|
|
\item \textbf{Coordinated attack}: 10 colluding workers attempting to submit false attestations were detected within 4 epochs, with all 10 slashed.
|
|
\end{enumerate}
|
|
|