Skip to content

FAQ.md

Latest commit

 

History

History
99 lines (60 loc) · 3.95 KB

FAQ.md

File metadata and controls

99 lines (60 loc) · 3.95 KB
Raw

FAQ — GPU ML-DSA / Dilithium signature verification

Answers to questions people often search for when looking for fast post-quantum signature verify on GPU.

What is this project?

MLDSA-GPU Verify is a CUDA batch signature verifier for ML-DSA (NIST FIPS 204), implemented for Dilithium2 / ML-DSA-44 today. You pass a public key and many (message, signature) pairs; the GPU returns valid/invalid per signature.

This repo is the public showcase: benchmarks, evaluation DLL, API docs, and licensing — not open CUDA kernel source.

Search terms this project matches

  • GPU Dilithium verify / Dilithium2 verification
  • ML-DSA batch verify, ML-DSA-44
  • Post-quantum cryptography (PQC) signature verification on NVIDIA GPU
  • CUDA lattice signatures, NTT on GPU (full pipeline, not NTT-only)
  • Faster than liboqs / pq-crystals CPU AVX2 for large batches
  • Blockchain validator / node throughput for PQC signatures
  • FIPS 204 verify acceleration
  • SHAKE256 / Keccak on GPU (inside verify pipeline)

How is this different from “GPU NTT only”?

Many experiments speed up only the Number Theoretic Transform. Real ML-DSA verify also needs unpacking, SHAKE hashing, matrix operations, hint processing, and final checks.

This project runs the entire verify path on GPU (with public key cached once). That is why batch throughput reaches hundreds of thousands of verify/s on RTX 3080-class hardware.

How fast is it?

On RTX 3080, batch 4096, one public key (see BENCHMARKS.md):

Path Throughput
CPU AVX2 (same machine) ~9,000 verify/s
GPU full pipeline (+ upload) ~632,000 verify/s
GPU full pipeline (replay) ~729,000 verify/s

Correctness: bit-exact vs pq-crystals/dilithium reference.

vs liboqs, pq-crystals, NVIDIA cuPQC?

MLDSA-GPU liboqs / pq-crystals NVIDIA cuPQC-PK
Target Production batch verify DLL Reference / general PQC lib Low-level PQC primitives SDK
Full ML-DSA verify on GPU Yes (published benchmarks) CPU-focused Build-it-yourself; few public verify numbers
Eval binary on GitHub Yes Open source SDK

Which GPUs are supported?

Evaluation builds are fat binaries: sm_86 (RTX 30xx), sm_89 (RTX 40xx), sm_90 (Hopper), sm_120 (RTX 50xx) + forward PTX.

Requirements: Windows x64, NVIDIA GPU, CUDA 12.x runtime.

Can I use this for Bitcoin / Ethereum / NEAR?

The library verifies raw message bytes you provide. It does not parse chain-specific transaction formats out of the box. Your node or adapter must serialize/hash transactions per your protocol, then call batch verify.

Commercial integration help (chain adapters) is available under license.

Is the CUDA source public?

No. CUDA kernels are proprietary. You get:

  • Free evaluation: bench_ntt.exe, libmldsa_gpu_verify.dll, demo_verify.exe
  • Commercial license: production rights + CUDA source (typically under NDA)

Contact: @MaxVip124 or commercial license issue.

How do I try it?

  1. Releases — download all files
  2. Run quick_test.bat (seconds) or run_benchmark.bat (minutes)
  3. Report your GPU results via benchmark issue template

Related standards and projects