papers/lp-201-zap-universal-wire/sections/bench.tex

% sections/bench.tex -- bench projections per surface
\section{Bench projections}
\label{sec:bench}

All measured numbers in this section come from in-tree benchmarks,
not vendor marketing. Headline ratios reproduce on commodity Apple
silicon (M-class, $10$ performance cores, Go~1.26).

\subsection{Measured baseline (ZAP vs alternatives)}

From \texttt{lux/zap/README.md}~\cite{zapreadme}, core operations:

\begin{center}
\begin{tabular}{lrrr}
\toprule
\textbf{Op} & \textbf{ns/op} & \textbf{B/op} & \textbf{allocs/op} \\
\midrule
\texttt{BenchmarkZAPParse-10}        & 2.9  & 0 & 0 \\
\texttt{BenchmarkZAPBuild-10}        & 44.0 & 0 & 0 \\
\texttt{BenchmarkConsensusRound-10}  & 5.5  & 0 & 0 \\
\bottomrule
\end{tabular}
\end{center}

ZAP versus MCP JSON-RPC (tool calling, 20-tool orchestrator):

\begin{center}
\begin{tabular}{lrrr}
\toprule
\textbf{Metric} & \textbf{MCP JSON-RPC} & \textbf{ZAP} & \textbf{Improvement} \\
\midrule
Serialisation       & 5{,}579~ns/op  & 322~ns/op   & $\bm{17\times}$  \\
Memory per call     & 2{,}826~B      & 256~B       & $\bm{11\times}$  \\
Allocations per op  & 58             & 2           & $\bm{29\times}$  \\
Parse time          & $\sim$1{,}000~ns & 2.9~ns    & $\bm{345\times}$ \\
GC runs per 50K ops & 41             & 3           & $\bm{14\times}$  \\
\bottomrule
\end{tabular}
\end{center}

Full 100K-operation memory profile: $304.01$~MB$\to$$26.70$~MB total
allocated ($91.2\%$ reduction); $6{,}001{,}513$$\to$$200{,}003$ mallocs
($96.7\%$ reduction). These are reproducible on the bench in
\texttt{lux/zap/bench/main.go}.

\subsection{LP-023 P-Chain measured delta}

The \texttt{vms/platformvm/txs/zap\_native/} subtree benchmarks
report $7.50\times$ Parse geomean over \texttt{linearcodec}, single
allocation of $24$~B versus $3$--$7$ allocations of $120$--$536$~B,
per memory \texttt{LP-023 P-chain ZAP-native}~\cite{lp023zapnative}.
That is the empirical floor for projecting blockchain wire migrations
in this paper.

\subsection{Magnetar v1.1 strict-atom delta}

Independent confirmation from the threshold-signing stack: Magnetar
v1.1.0 strict-atom is measured $35$--$51\%$ faster than the circl
SHAKE-192s baseline at byte identity, per memory
\texttt{magnetar\_v110\_strict\_atom}~\cite{magnetar}. The
mechanism is the same: zero-allocation segment-pointer decode
replacing reflect-driven copy.

\subsection{Projected per-surface delta}

We project the LP-023 $7.50\times$ Parse geomean onto every Tier-1
blockchain-wire surface (P-Chain, X-Chain, UTXO, warp message,
proposerVM, chain VMs). For each AI/ML Tier-1 surface we project the
ZAP-vs-JSON-RPC $17\times$ serialisation / $345\times$ parse / $96.7\%$
allocation ratio.

\begin{center}
\begin{tabular}{lll}
\toprule
\textbf{Surface (Tier 1)} & \textbf{Current op cost} & \textbf{Projected} \\
\midrule
P-Chain tx parse        & $\sim$$1.0$~\textmu s & $\sim$$130$~ns ($7.5\times$) \\
X-Chain UTXO parse      & $\sim$$0.8$~\textmu s & $\sim$$105$~ns ($7.5\times$) \\
Warp message verify     & $\sim$$2.3$~\textmu s & $\sim$$300$~ns ($7.5\times$) \\
Agent tool call ser     & $5{,}579$~ns          & $322$~ns ($17\times$) \\
Agent tool call parse   & $\sim$$1{,}000$~ns    & $2.9$~ns ($345\times$) \\
MCP per-call alloc      & $58$ allocs           & $2$ allocs ($29\times$) \\
\bottomrule
\end{tabular}
\end{center}

\paragraph{Stack-wide compounding.}
On a hypothetical RPC-heavy synthetic transaction~--- one P-Chain tx
verify, two X-Chain UTXO parses, one warp message verify, four agent
tool calls, two MCP round-trips~--- the current end-to-end CPU cost
sums to roughly $26$--$30$~\textmu s. The projection sums to roughly
$3.5$~\textmu s, a $\bm{7.4\text{--}8.6\times}$ compounded speedup
across the AI/ML/blockchain path. The allocation reduction is more
extreme: roughly $238$~allocations~$\to$~$\sim$$10$~allocations, a
$\bm{24\times}$ reduction translating directly to GC pressure
relief and tail-latency floor improvement.

\paragraph{Caveats.}
These projections assume each surface's wire cost dominates its own
op latency. For surfaces where wire is a minority of total work (e.g.
ML-DSA verification at $\sim$$600$~\textmu s end-to-end where wire is
$\sim$$5\%$), the projected speedup is dampened by Amdahl. Wire-bound
hot paths (block parsing, tool-call orchestration, tensor-stream
demux) realise the full ratio.

\paragraph{Confidence.}
Strong on (1)--(4): same code shape as already-measured surfaces.
Weak on Jin tensor wire: schema is not yet specced. Listed as
\emph{projection pending schema}.