Releases
v2.0.0
Compare
Sorry, something went wrong.
No results found
Performance
~410M keys/sec on RTX 5080 (v1.0.0: 67-70M on the same card).
Fixed-base comb with W-bit signed windows (W = 8..18, default from L2 size, --tune to measure);
window 0 folded into the start point.
GF(2^255-19) on 8 x 32-bit limbs with PTX carry chains, Karatsuba multiply (tools/gen_fe8.py).
Batch inversion moved to its own kernel (128 keys per inversion), overlapped with key generation on a
high-priority stream.
L2 prefetch of the next table entry.
Features
Prefix (-p), suffix (-s) and prefix+suffix (-m P*S) patterns, several at once.
Case-insensitive matching (-i).
Exact odds and expected time per match.
Multi-GPU (-d 0,1 / -d all).
--selftest, --tune, --bench, --profile.
The old vanity_gpu <suffix> [count] form still works.
Security
Seeds from ChaCha12 keyed with the OS RNG instead of rand()/time(). Keys made by v1.0.0 should
be considered weak.
Every match is recomputed on the CPU with an independent reference implementation before it is written.
Fixes
Suffixes longer than 10 characters no longer overflow silently.
Speed is measured with a wall clock on every platform.
Matches beyond the per-batch buffer are counted and reported instead of being lost silently.
License
Apache License 2.0 with a NOTICE file (was MIT): redistributions must keep the author credit.
Build
build.bat / build.sh build for the local GPU by default (SM_ARCH overrides).
Release packages add sm_120 (RTX 50xx); CI uses CUDA 12.9.
You can’t perform that action at this time.