GPU Acceleration

Performance

NOTE: as of 5/11 these numbers are out of date

Measured throughput on real hardware:

	Throughput	Hardware
CPU	~1.3M keys/sec	AMD Ryzen 9 7950X3D (32 threads)
GPU	~68M keys/sec	NVIDIA RTX 4090 (24GB)

The GPU is roughly 50x faster than the CPU on this hardware. The speedup comes from running the full Ed25519 pipeline across thousands of GPU threads in parallel.

Building with CUDA

cargo build --release --features cuda

Requires the NVIDIA CUDA Toolkit to be installed. The CUDA kernel is compiled at runtime via NVRTC, so nvrtc and cuda shared libraries must be on PATH.

The cudarc dependency is configured for CUDA 13.1 by default. To target a different CUDA version, change the cuda-13010 feature in Cargo.toml to match your installation (e.g. cuda-12060 for CUDA 12.6).

How GPU mode works

When --gpu-only (or hybrid) is selected, the entire keygen pipeline runs on the GPU. A single host thread loops: launch kernel batch, sync, check for match, repeat. The TUI progress display works identically for both CPU and GPU modes.

Each GpuSearcher draws a fresh 32-byte start_scalar from the OS CSPRNG and advances it by 8 * keys_checked between batches. Per launch, every thread runs one ge_scalarmult_base on start_scalar + 8 * tid to derive its starting point A, then iterates ITERS_PER_THREAD = 256 cheap point additions of 8B (using base[0][7] = 8B as a ge_precomp) — each iteration produces a fresh candidate pubkey from A and checks the prefix. To amortize the expensive fe_invert needed for point compression, the kernel uses Montgomery batch inversion with B = 16: one inversion per 16 points instead of one per point. On every prefix hit the kernel writes back the scalar and pubkey, and the host runs a defensive scalar·B == pubkey check via curve25519-dalek before accepting the match.

The Ed25519 field arithmetic uses the ref10 implementation (radix-$2^{25}$.5, int32[10] limbs) from solana-perf-libs, which is a CUDA adaptation of the SUPERCOP reference code. Scalar multiplication uses windowed lookup tables with 32 precomputed basepoint multiples.

Verification

Use --verify to cross-check GPU-generated keys against the CPU implementation. This is useful for validating correctness after modifying the CUDA kernel.

Use --benchmark <SECS> to run a separate vanity_count_matches kernel that never exits early on a hit. It atomically tallies every match over the given duration and compares the observed count against the expected count implied by the reported keys/sec rate — a sanity check that the reported throughput isn't inflated by mid-launch early exits. Works on both the CUDA and Metal backends.

Sources

• solana-perf-libs — CUDA Ed25519 field arithmetic (ref10)
• cudarc — safe Rust bindings for the CUDA driver API