Performance measurements

Performance figures describe a measured implementation on a named machine. They are not consensus constants and should not be projected onto an unrelated CPU from core count alone.

The reference system used:

CPU       Intel Core i7-1365U
Topology  10 cores / 12 threads
ISA       AVX2 + VPCLMULQDQ, no AVX-512
OS        Linux 6.17 x86-64
Rust      1.96.0

Each table reports 20 measured samples after warm-up. p95 uses the nearest-rank estimator.

Wallet authorization path

The wallet benchmark includes page construction, logical hashing, one authorization capsule, complete intent encode/decode and local capsule admission. It excludes network latency and block HistoryStep proving.

Case	Pages	Total p50	Total p95	Proof / intent size
1 input	1	228.30 ms	352.47 ms	56.49 / 56.81 KiB
100 inputs	13	217.32 ms	255.46 ms	56.58 / 60.69 KiB
1,020 inputs	128	233.06 ms	285.81 ms	56.11 / 96.50 KiB

The 1,020-input case still produces one authorization capsule. Larger input sets increase page hashing and serialized intent size, but not the number of wallet proofs.

Reproduce the harness with:

NOID_WALLET_BENCH_SAMPLES=20 cargo run --release \
  --manifest-path research/two_class/Cargo.toml \
  --bin two-class-wallet-bench

HistoryStep classes

Complete preparation is assembly plus proving. Verification measures the terminal verifier used by a full node.

Class	Useful rows	Prepare p50 / p95	Verify p50 / p95	Terminal
B64, `m=23`	5,705,307	11.472 / 14.387 s	0.666 / 0.720 s	766,549 B
B255, `m=24`	15,368,233	24.189 / 29.755 s	0.770 / 1.012 s	807,189 B

On this host, B64 passed the strict 15-second p95 preparation gate with 613 ms of margin. B255 did not, so the production capacity selector correctly kept the miner at B64. Both classes remain inexpensive for an ordinary node to verify relative to proving them.

The benchmark used Thin LTO, one codegen unit and target-cpu=native to measure the exact host. Official binaries keep a portable process baseline and select proof and PoW kernels at runtime; therefore release-package timing must also be checked on its destination.

Reading the figures

The wallet table measures local authorization, not time to confirmation. The HistoryStep table measures block preparation, not expected proof-of-work search time. Network propagation and nonce search vary independently.

For a miner, qualify the complete path:

select intents
  + assemble witness
  + prove HistoryStep
  + search nonce
  + submit and accept block

Optimizing one internal phase does not justify enabling a larger proof class if complete preparation misses the target interval.

The archived reports in the source repository record command lines, commits, matrix digests and raw samples under research/two_class/results/.