Project: High GPU Throughput Bitcoin secpk256 cracking

[ipsbruno3]

secp256k1 Sequential Scan — GPU Throughput (with working math)

Hello, I built an OpenCL implementation of secp256k1 that can process millions to billions per second, generate a public keys on a modern GPU.

https://github.com/ipsbruno3/secp256k1-gpu-accelerator

TL;DR

• Per-GPU (RTX 5090): ~500–600M public keys/second minimum
• 12× rig: ~7.2–10.0B public keys/second observed
• Per day: with 600M×12 → $$6.2208\times 10^{14}$$ keys/day; with 10B×12 → $$8.64\times 10^{14}$$ keys/day
• wNAF: Larger windows improved scalar-mult throughput (trade VRAM for fewer additions)

To-do

• Used-address hash set: O(1) membership checks; increases detection, not raw throughput

---

Throughput math

Let:

• $$R_{\mathrm{gpu}}$$ = per-GPU public keys per second
• $$G$$ = number of GPUs
• $$T_{\mathrm{day}} = 86{,}400\ \mathrm{s}$$
• $$R_{\mathrm{rig}} = R_{\mathrm{gpu}}\cdot G$$
• $$K_{\mathrm{day}} = R_{\mathrm{rig}}\cdot T_{\mathrm{day}}$$

Example (600M keys/s per GPU, 12 GPUs):

$$ R_{\mathrm{gpu}} = 6.0\times 10^{8}\ \frac{\text{keys}}{\text{s}},\quad G=12,\quad T_{\mathrm{day}}=86{,}400\ \text{s} $$

$$ R_{\mathrm{rig}} = 6.0\times 10^{8}\cdot 12 = 7.2\times 10^{9}\ \frac{\text{keys}}{\text{s}} $$

$$ K_{\mathrm{day}} = 7.2\times 10^{9}\cdot 86{,}400 = 6.2208\times 10^{14}\ \text{keys/day} $$

per day

And if you multiplier with hashtable used iaddress (53 millions)

per day validations (don’t multiply by set size)

---

Hit probability (order-of-magnitude)

With $$N \approx 5.3\times 10^{7}$$ known used addresses and address space $$M=2^{160}$$:

$$ \mathbb{E}[\text{hits/day}] \approx K_{\mathrm{day}}\cdot \frac{N}{M}. $$

Using $$K_{\mathrm{day}}=6.2208\times 10^{14}$$:

$$ \mathbb{E}\approx \frac{6.2208\times 10^{14}\cdot 5.3\times 10^{7}}{2^{160}} \approx 2.26\times 10^{-26}\ \text{hits/day}. $$

Brute force on random keys stays infeasible. You only get traction with constrained keyspaces (partial seeds, weak RNGs, human patterns, etc.).

---

Implementation notes

• wNAF window: Larger $$w$$ reduced additions and improved throughput on 5090s; optimal $$w$$ depends on VRAM vs. occupancy.
• Used-address filter: Keep it as an in-memory hash set; serialize once and memory-map for fast cold starts.
• I/O: Batch keys → compress/pack → single pass membership checks to avoid cache thrash.
• Ethics: Test only against keys you own.

---

Other projects

https://github.com/ipsbrunoreserva/bitcoin_cracking
— High-performance PBKDF2-HMAC-SHA512 (OpenCL). WIP.

https://github.com/ipsbruno3/bitcoin_cracking_final
— Continuation of @ipsbrunoreserva/bitcoin_cracking_final. WIP (Final Version).

https://github.com/ipsbruno3/bitcoin_electrum_cracking
— Electrum seed verification trick + sequential BIP-39 scan + used-address membership checks (Works 100%)

---