\begin{table}[htbp]\centering\small
\caption{Measured training throughput and memory on the compute node (Apple M3 Ultra{,} m3u96b{,} torch 2.15.0{,} mps; FP32)}
\label{tab:benchmark}
\begin{tabular}{lrrrrrr}
\toprule
Preset & Batch & Parameters & Steps/s & Examples/s & Peak memory (MB) & Latency/origin (ms) \\
\midrule
small & 32 & 3{,}455{,}619 & 20.81 & 666 & 54 & 0.34 \\
small & 64 & 3{,}455{,}619 & 12.21 & 782 & 69 & 0.25 \\
small & 128 & 3{,}455{,}619 & 5.93 & 758 & 97 & 0.27 \\
spec & 32 & 7{,}677{,}955 & 11.54 & 369 & 114 & 0.59 \\
spec & 64 & 7{,}677{,}955 & 6.29 & 402 & 119 & 0.48 \\
spec & 128 & 7{,}677{,}955 & 2.97 & 380 & 147 & 0.53 \\
base & 32 & 20{,}718{,}339 & 6.11 & 195 & 275 & 5.04 \\
base & 64 & 20{,}718{,}339 & 3.24 & 207 & 284 & 1.59 \\
base & 128 & 20{,}718{,}339 & 1.51 & 193 & 315 & 1.76 \\
\midrule
Data loader (64/batch) & 0 workers: 3971 ex/s; 4 workers: 13016 ex/s; 8 workers: 38486 ex/s & & & & & \\
\bottomrule
\end{tabular}
\end{table}