Historical performance — v1 / v1.1

This page preserves measurements from the legacy profiles. Current v2 measurements are in Performance.

Performance belongs to one source revision, proof profile, authenticated matrix pack, build profile and host. It is not a consensus constant and cannot be inferred from core count alone.

The historical construction measurements below use Parano1d revision 39626b22d53cf2f2c480a7e28446c197dca68043, the production C1 profile and the authenticated B25/B255 matrix pack. They predate v1.1 shared-path encoding. They remain a hardware baseline for later capacity experiments, including v2; new proof shapes and complete block production still need their own measurements.

The original measurement record preserves the expanded terminal sizes from that revision. Those sizes are not the v1.1 network payload, and these timings exclude shared-path codec overhead.

PoW nonce search is not included in the table. ASERT targets the complete elapsed interval between accepted blocks. It does not assign a separate 20-second budget to nonce search. Proof preparation, nonce search and network propagation all occupy the same observed block interval, and ASERT adjusts the nonce target against that complete cadence.

Terminal size in v1.1

Shared-path encoding stores and transmits each shared authentication node once. A fresh codec audit on 2026-09-24 measured the same verified proof in both representations, using the production C1 profile and authenticated B25/B255 pack:

Both samples are below 1 MB (1,000,000 bytes). These are one-proof examples, not fixed sizes or upper bounds: shared-path size depends on query openings. An independently verified mainnet B25 terminal at height 137191 was 872,500 B. The consensus cap remains 1,100,000 bytes for encoded and expanded terminals. Block bodies, transport framing and RPC hex text are excluded from these sizes.

The codec measurement record contains reproduction instructions and validation results. This audit measures representation savings; it does not remeasure the historical AVX-512 host.

Native state processing

The node retains a bounded cache of authenticated segment columns and their exact Merkle trees. The payload budget is 64 MiB per state view, including the columns: nine production-size segments fit. Copies share immutable data until modified. Active block scratch is separate from this retained-cache budget.

Hot slot updates recalculate changed paths. Cold loads authenticate the complete segment against its exact root before entering the cache. After commit, cold payloads are evicted; restart and snapshot installation rebuild the cache on demand. The same policy applies across the whole slot domain. Workloads with poor locality still pay for cold authentication, so measure them separately from repeated updates to one hot segment.

The ignored state benchmark covers a dense segment, eight and sixteen touched segments, and repeated misses across thirty-two segments. Each case compares its final root with the streaming reference. It measures authenticated loads and native root updates; disk commits, HistoryStep verification and PoW are outside the measured intervals. The first iteration starts cold.

RAYON_NUM_THREADS=4 cargo test --locked --release -p noid_chain --lib \
  bench_exact_state_cache_cycles -- --ignored --nocapture --test-threads=1

For comparisons, fix CPU affinity, backend and build profile for both binaries. Keep a separate result for each workload and for cold versus warmed iterations.

Wallet authorization

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

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

The production C1 wallet uses 65 Fiat–Shamir queries. One PagedSpend contains one authorization capsule whether it occupies one page or the full 128 pages. The canonical serialized authorization has a 92,696-byte worst-case bound.

HistoryStep

The isolated production benchmark requires a completed and authenticated matrix pack. Run each class separately so the output identifies the exact parent and child class.

NOID_PACK_ROOT=../parano1d-artifacts/history-step-pack-v1
source "$NOID_PACK_ROOT/pins.env"
export NOID_HISTORY_STEP_RUNTIME_METADATA_RELEASE_DIGEST
export NOID_HISTORY_STEP_PACK_LEAF_DIGESTS
export NOID_HISTORY_STEP_PACK_DIR="$NOID_PACK_ROOT"

NOID_HISTORY_STEP_BENCH_FILTER=B25 \
NOID_HISTORY_STEP_BENCH_SAMPLES=20 \
cargo bench --locked -p bench_prover --bench history_step_proof

NOID_HISTORY_STEP_BENCH_FILTER=B255 \
NOID_HISTORY_STEP_BENCH_SAMPLES=20 \
cargo bench --locked -p bench_prover --bench history_step_proof

cargo bench uses the optimized bench profile. Transaction construction, wallet proving, block-template construction and matrix authentication are setup. history_step_ms covers parent-terminal decoding, bounded input and authorization preparation, recursive assembly, nonce sealing, proof construction and terminal encoding. verify_ms covers bounded wire decoding and complete terminal verification.

End-to-end block production

The isolated proof measurement is not the complete mining latency. Capacity decisions must measure:

select intents
  + assemble the current block trace
  + replay and bind the parent terminal
  + prove HistoryStep
  + search the nonce
  + submit and accept the block

Nonce search and network propagation vary independently from proof construction. End-to-end comparisons must use the complete production path on the final host. The miner's automatic B255 permission uses only the first completed B25 preparation and a four-times timing estimate, as described in Mining architecture. Official binaries keep a portable baseline and select the pclmul, avx2+vpclmul, avx512bw+vpclmul or neon+pmull backend at runtime.