Batch Modular Inversion for ECDLP: A Case Study with FastscanGPU

[ethicbrudhack]

Hi all,

I've been exploring optimizations for the Elliptic Curve Discrete Logarithm Problem (ECDLP) on GPUs and wanted to share a practical implementation I've been working on. It's called FastscanGPU.

One of the key techniques we used is batch modular inversion to speed up point addition on the secp256k1 curve. Combined with a Bloom filter for efficient address matching, we've managed to achieve around 3.5 GH/s on a single RTX 4090.

The project is open-source and might be useful if you're working on similar performance optimizations:
https://github.com/ethicbrudhack/FastscanGPU-BitcoinPuzzleChallengeBreaker-

I'm particularly interested in any feedback on the inversion approach. Has anyone experimented with similar techniques for multi-GPU setups?

Thanks!

[solardiz]

Thanks, but I don't immediately see where this would benefit our project - do you?

We currently have uses of secp256k1 in 3 formats - for Armory, BitShares, and Electrum wallets. Out of these, Armory and Electrum have such slow KDFs that secp256k1 performance almost does not matter. There's no potential for GH/s figures like yours anyway - it's more like kH/s.

For BitShares, it could matter, but our current implementation is mostly non-optimized in many other aspects as well. It's CPU-only without SIMD, and no one seems to care - maybe those wallets are too obscure.