Wallet password recovery benchmarks — measured, reproducible, dated.

The hardware and how it was tested

The numbers come from the official btcrecover benchmark suite (measured 2026-07-04), which publishes system, tool and per-wallet results. The systems that produced the figures used on this site:

System #2 is the reference used for the rounded whole numbers on this site (e.g. 472, 11,950, 5,160). Where a row exists for CPU and GPU, both are shown.

Measured password rates (guesses per second)

Rounded from the published runs (9950X: 472.17 / 122.12 / 72.01 / 15.25K / 627.58K / 2.77K / 5.45K / 503.42 / 29.67K / 95.72 / 5.16K / 1.81M / 509.59; RTX 5090: 11.95K / 3.21K / 42.88 / 19.95K / 9.47K / 20.03K / 4.30K / 5.59K / 529.40K). GPU cells marked “—” are memory-hard formats where OpenCL was not run or barely helps. The Ethereum row is the suite’s reference test (N=2¹⁷ as published); the geth/Web3 standard is N=2¹⁸ = 262,144, which tests roughly 2× slower per guess on the same hardware.

The MetaMask row is the legacy 10,000-iteration vault. Newer vaults store their own KDF parameters in the vault data — read the actual count (600,000 and 900,000 are both documented) — which is estimated at up to roughly 90× slower per guess, an extrapolation from the iteration ratio, not a separate 900k benchmark run. The same tools and commands apply, with a much longer wait.

The BIP39 passphrase row has both a CPU and a GPU/OpenCL measurement in the official benchmark set (RTX 5090 ≈ 19,950 guesses/sec); use whichever matches your hardware.

Two rows look odd until you know the cause: the MultiBit Classic .key and Electrum 2.8+ passphrase rows show GPU rates below the CPU rate. That is not a typo — the OpenCL path for those two formats in btcrecover is slower than the optimized CPU path, so the CPU column is the faster choice for them.

Why ranges beat single numbers

Real-world speeds vary with the tool version, the GPU driver, background load and — most of all — the KDF parameters inside your specific file. A wallet.dat encrypted on a 2013 laptop can test 4× faster than one encrypted on a 2024 desktop, because the iteration count is stored in the file, not fixed by the format. That is why this site says “hundreds” or “low thousands” in prose and points here for the measured figures, rather than pretending one number is your number.

Use the strength calculator to combine a speed class with your own search space, and the cracking-time guide for what the result means in hours, years or centuries.

How to reproduce or extend these numbers

1. 1

   Run the suite yourself

   The btcrecover project publishes the exact test commands with the benchmark page; the same repository also documents GPU acceleration setup and version pinning.

2. 2

   Record the full context

   Hardware model, tool version, driver, wallet file or test vector, KDF parameters, date. A number without that context is not a benchmark.

3. 3

   Report ranges, not records

   One fast run on a cold machine is a fluke. When benchmarking your own hardware, repeat the test several times to estimate normal variance rather than treating the fastest single run as definitive.

Freshness policy

These figures are reviewed quarterly and re-measured when a major tool or GPU generation changes the picture. Every speed table on this site links back to this page, so a reader can always see how old the number is and on what hardware it was produced.

Last reviewed: 14 August 2026 · Next scheduled review: November 2026.

Sources & references

1. BTCRecover — official performance benchmarks (measured 2026-07-04)
2. BTCRecover — GPU acceleration documentation