Skip to content

SPEEDS.md

Latest commit

 

History

History
78 lines (60 loc) · 4.1 KB

SPEEDS.md

File metadata and controls

78 lines (60 loc) · 4.1 KB
Raw

Measured speeds (TrueCollider / KeyCollider)

Real rates captured from live binaries on the maintainer benchmark host.
Re-run: powershell -ExecutionPolicy Bypass -File .\scripts\run_benchmarks.ps1

Hardware

Component Spec
CPU Intel Core i7-920 @ 2.67 GHz · 8 logical threads · SSSE3 (SSE hash160 path)
GPU NVIDIA GeForce RTX 3060 Ti 8 GB
Builds keyhunt.exe (MinGW) · keyhunt_cuda.exe (VS2022 + CUDA 12.8)
Date 2026-07-13

Method

  • Stats interval -s 5, sample 15 s sustained (and 5 s peak for CUDA).
  • Random walk in a large hex range so the run does not finish early.
  • Rates taken from the last [+] Total ... (N keys/s) line printed by the tool.

CPU results

Mode Example command 15 s rate
address BTC keyhunt -m address -f targets.txt -l compress -t 8 -A sse -q -s 5 6,561,245 keys/s
address BTC + endo add -e 8,163,328 keys/s
address ETH -c eth -t 8 3,240,413 keys/s
address SOL -c sol -t 8 70,724 keys/s
rmd160 -m rmd160 -t 8 7,325,559 keys/s
rmd160 + endo add -e 8,102,707 keys/s
xpoint -m xpoint -t 8 11,329,467 keys/s
vanity -m vanity -v 1Love -e -t 8 8,067,481 keys/s
pubkey2addr -m pubkey2addr -t 8 489,009,152 keys/s*
minikeys -m minikeys -t 8 48,537 keys/s
mnemonic -m mnemonic -t 8 247,057 mnemonics/s
poetry -m poetry -t 4 57,862,758 mnemonics/s
brainwallet -m brainwallet -t 4 94,648,934 mnemonics/s
bsgs -m bsgs -b 40 -n 1048576 -t 8 26,442,255,564 keys/s†

* As reported by the binary; not necessarily comparable 1:1 with address.
† Effective BSGS coverage rate with a 1M baby table.

CUDA results (RTX 3060 Ti) — custom CUDA edition

Mode Command Sustained / peak
address BTC (sequential GRP=1024) -U cuda -M auto -t 1 -x sequential ~90–96 Mkeys/s peak (large auto batch)
address BTC (prior GRP=256) same ~165–166 Mkeys/s (more centers / occupancy)
rmd160 (sequential GRP) same ~same EC+hash path
xpoint (sequential GRP) -m xpoint -U cuda … higher than address when GRP path active
address ETH -c eth -U cuda -M auto -t 1 lower (Keccak often host-side)
mnemonic -m mnemonic -U cuda mnemonics/s (PBKDF2-bound; not EC Mk/s)

Path: device GRP keystream (G-table + Montgomery batch inv, GRP=1024, PreferL1, auto host batch up to ~8–14M keys so thousands of centers stay in flight). Falls back to per-key scalar_mul_g for non-consecutive keys, ETH, endomorphism, or host-filter mode. Prefer -M auto / -t 1.

KeyHunt-Cuda README ~1013 Mk/s is XPOINT on a GTX 1650 with GRP=2048 — not 1:1 with BTC address+hash160 on a 3060 Ti. GRP=1024 trades some occupancy vs GRP=256 on this card; further gain needs less host packing overhead / dual streams.

Measured 2026-07-26: GRP=256 ≈165–166 Mkeys/s; GRP=1024 + larger batches ≈90–96 Mkeys/s peak (was ~15–16 with per-key scalar×G; ~24 with GRP=1024 undersubscribed).

GPU EC wired vs still CPU-only

Mode GPU?
address / rmd160 / ETH / troot / vanity / xpoint / pubkey2addr / minikeys Yes — GPU EC; BTC-family uses device hash160+bloom when ready
mnemonic / poetry / brainwallet Yes — GPU EC after derive
bsgs Yes — GPU baby-table + device GRP (host bloom; serial cycles)
address -c sol Yes — full device ed25519 ge
Full on-device hash160+bloom Yes when self-test passes
kangaroo Yes — CUDA batch EC scan (≤2²⁴) + multi-walker DP; CPU fallback

See hub README.md for commands and BSGS -n/-k tables.

Notes

  • AVX2/AVX-512 CPUs should beat these SSE-only CPU figures substantially for hash160 modes.
  • Sequential CUDA GRP (~165 Mkeys/s) now beats this host's SSE CPU address path; sparse/non-stride-1 modes still use the older per-key GPU EC path.
  • Prefer -t 1 with CUDA to avoid GPU lock contention.
  • Roadmap leftovers (Kangaroo GPU, BSGS GRP throughput): ROADMAP.md.