SIMD-0565 gets proposed compute-unit prices for secp256r1 syscalls
Victoria Shinder2 min readA pull request opened on 14 October 2026 in the Solana Improvement Documents repository, #695, adds a "Compute Unit Usage" section to SIMD-0565, the proposal for secp256r1 syscalls. The author, samkim-crypto, proposes concrete compute-unit (CU) prices for each operation. The pull request is open, not merged.
The proposed prices.
- point validation: 20 CU
- decompression: 610 CU
- addition and subtraction: 520 CU each
- scalar multiplication: 9,000 CU
- multi-scalar multiplication (MSM): 9,000 + 3,800 × (points − 1) CU
How they were set. The author calibrated the numbers on a private cluster of ten equally staked validators — nine AMD EPYC 9354P and one EPYC 9354 machine, 32 cores each and 755 GB of RAM — running an Agave-based client (4.4.0-alpha.5) with the syscalls ported in, Alpenglow with 200 ms slots and the SIMD-0286 limits of 20 million CU per writable account and 50 million per block. A leader filled blocks with transactions that all write one shared account, forcing followers to replay them serially up to the per-account limit. Starting from microbenchmark estimates, each operation's price was raised until replay times returned to the level of empty slots, which already take about 200 ms.
The criterion. The target is stated in terms of finality: the follower that completes 80% of stake — the seventh of nine, which gates fast finalisation — should replay saturated blocks about as fast as empty ones. Scalar multiplication settled at 9,000 CU for 64 µs of execution, about 7.1 ns per CU, and the other operations were priced at the same rate from their measured per-call times. With the proposed costs, the seventh follower's worst-slot median replay for 64 scalar multiplications per transaction fell from 229.8 ms at the starting price to 202.9 ms, against 199.7 ms for empty slots; the slowest follower ran about 6 ms over. Point validation stayed at 20 CU because 1,024 calls per transaction already replayed at empty-slot speed.
What it shows beyond this one SIMD. The write-up also tested transactions that write separate accounts and so replay in parallel up to the block limit; there replay stayed close to empty-slot level at every price tested. The shared-account case is what sets the price — a useful reminder that on Solana a syscall is priced for the worst serial path, not for average throughput.
Source: Solana Improvement Documents, pull request #695 "[SIMD-0565] Specify compute unit costs for the SIMD-0565 secp256r1 syscalls" — https://github.com/solana-foundation/solana-improvement-documents/pull/695