Execution Proof Performance Report

RM Execution proof
1

Execution Proof Guest — Performance Report

Benchmark Results (wall time, Groth16 prover)

  • with delta — delta proof verified inside the zkVM via ECDSA recovery
  • without delta — delta proof bytes passed through to the journal; ECDSA verification is skipped in the circuit and delegated to an external verifier

tx_count (1 compliance unit per transaction)

tx_count with delta (median) without delta (median) Δ
1 8.520 s 6.759 s −20.7%
2 10.328 s 7.193 s −30.4%
4 15.304 s 8.442 s −44.8%
8 25.526 s 11.751 s −54.0%
16 45.412 s 18.498 s −59.3%

compliance_count (1 transaction, N compliance units)

compliance_count with delta (median) without delta (median) Δ
1 8.546 s 6.811 s −20.3%
2 9.062 s 6.806 s −24.9%
4 9.678 s 7.318 s −24.4%

Cycle Profile

Two profiles are tracked:

  • profile.pb — baseline: delta proof verified inside the zkVM (ECDSA recovery + point check)
  • profile_without_delta.pb — delta proof passed through to the journal; external verifier checks the signature
profile.pb profile_without_delta.pb
Total cycles 1,056,572 285,165 (−73.0%)
Accounted cycles 1,034,990 (97.96%) 283,672 (99.48%)

profile.pb — Top Functions by Self Cycles

Rank Function Flat cycles Flat % Cum %
1 sys_bigint 274,647 25.99% 26.12%
2 sys_bigint2_3 167,444 15.85% 15.91%
3 [PageIn] 92,650 8.77% 8.77%
4 sys_bigint2_4 90,447 8.56% 8.59%
5 FieldElement::invert 59,670 5.65% 25.04%
6 sys_read_words 48,776 4.62% 4.65%
7 AffinePoint::mul 45,388 4.30% 30.99%
8 memcpy 44,915 4.25% 5.58%
9 memcmp 21,853 2.07% 2.24%
10 Vec<u32>::write_words 20,249 1.92% 4.04%
11 sys_write 20,064 1.90% 2.21%
12 [PageOut] 17,201 1.63% 1.63%
13 keccak::keccak_p 16,160 1.53% 1.72%
14 sys_sha_buffer 15,630 1.48% 2.19%
15 Scalar::invert 14,738 1.39% 6.80%
16 FdWriter::write_words 14,670 1.39% 5.57%
17 FdReader::read_words 10,318 0.98% 5.66%
18 VecVisitor::visit_seq 7,693 0.73% 7.38%
19 FieldElement::sqrt 6,623 0.63% 3.03%
20 ExpirableBlob::serialize 5,704 0.54% 5.96%

profile.pb — Key Call-Chain Hotspots

Call Chain Cum Cycles Cum %
execution_proof_guest::main 1,036,034 98.06%
DeltaProof::verify 655,789 62.07%
ecdsa::recover_from_digest 653,539 61.85%
k256::mul::lincomb 450,931 42.68%
AffinePoint::mul 327,425 30.99%
ecdsa::hazmat::verify_prehashed 287,496 27.21%
ProjectivePoint::to_affine 208,716 19.75%
DeltaInstance::from_deltas 97,169 9.20%
deserialize_struct (serde) 87,142 8.25%
InsertionWitness::apply 28,201 2.67%

profile.pb — Analysis

The circuit is dominated by ECDSA (DeltaProof::verify at 62.07% cumulative):

Category Cycles Share
ECDSA / delta proof (sys_bigint, sys_bigint2_3/4, FieldElement::invert, AffinePoint::mul) ~580K ~55% flat
Delta instance (DeltaInstance::from_deltas) ~97K ~9% cum
Memory paging ([PageIn], [PageOut]) ~110K ~10%
Serde I/O (read + write path) ~100K ~9%
Tree ops (InsertionWitness::apply) ~28K ~3%

The ECDSA cost is near-floor for secp256k1 on RISC0 — sys_bigint and sys_bigint2_3/4 are already hardware-accelerated syscalls. The only way to materially reduce this is to eliminate the verification from the circuit entirely (as done in profile_without_delta.pb) or to aggregate multiple delta proofs into a single signature.

profile_without_delta.pb — Top Functions by Self Cycles

Rank Function Flat cycles Flat % Cum %
1 sys_read_words 48,776 17.10% 17.10%
2 [PageIn] 42,860 15.03% 32.13%
3 memcpy 34,350 12.05% 44.18%
4 sys_write 24,420 8.56% 52.74%
5 Vec<u32>::write_words 20,249 7.10% 59.84%
6 FdWriter::write_words 17,852 6.26% 66.10%
7 sys_sha_buffer 16,086 5.64% 71.75%
8 [PageOut] 10,339 3.63% 75.37%
9 FdReader::read_words 10,318 3.62% 78.99%
10 VecVisitor::visit_seq 7,693 2.70% 81.69%
11 ExpirableBlob::serialize 5,704 2.00% 83.69%
12 sha::copy_and_update 5,264 1.85% 85.53%
13 to_vec (serde) 4,089 1.43% 88.66%
14 deserialize_tuple 3,212 1.13% 91.13%
15 deserialize_struct 3,138 1.10% 93.33%
16 InsertionWitness::apply 615 0.22% 98.33%

profile_without_delta.pb — Key Call-Chain Hotspots

Call Chain Cum Cycles Cum %
execution_proof_guest::main 269,746 94.59%
ExecutionProofInstance::serialize 63,914 22.41%
TxInfo::serialize 55,373 19.42%
AppData::serialize 55,274 19.38%
deserialize_struct (serde) 88,394 31.00%
VecVisitor::visit_seq 78,310 27.46%
InsertionWitness::apply 28,832 10.11%
to_vec (serde) 29,675 10.41%
env::verify 14,235 4.99%

Analysis

Without ECDSA, the circuit is dominated by serde I/O:

Category Cycles Share
Read path (sys_read_words, FdReader, VecVisitor) ~66,787 ~23%
Write path (sys_write, Vec::write_words, FdWriter, ExpirableBlob::serialize, AppData::serialize) ~123,478 ~43%
Memory paging ([PageIn], [PageOut]) ~53,199 ~19%
Tree ops (InsertionWitness::apply, IncrementalMerkleTree::insert) ~31,016 ~11%

The write path is now the single largest cost at ~43%, driven almost entirely by AppData / ExpirableBlob serialisation in env::commit. The TxInfo::serialize cumulative (19.42%) and AppData::serialize cumulative (19.38%) confirm these are essentially the same call chain.

Version with delta: GitHub - anoma/arm-risc0 at xuyang/execution_proof · GitHub
Version without delta: GitHub - anoma/arm-risc0 at xuyang/execution_proof_without_delta · GitHub

2 Likes