BENCHMARK

mdxfind documentation

Benchmarks

Standard Benchmark Test Set

A reproducible benchmark suite is available for download from www.mdxfind.com:

File Contents Size
rockyou.txt.gz Rockyou wordlist (14.3M passwords) 49MB
mdxfind-benchmark-full.zip Full hash test files (14.3M hashes each, unsalted + salted) 1.7GB
mdxfind-benchmark-small.zip Small hash test files (1M hashes each, unsalted + salted) 122MB
mdxfind-benchmark-2811.zip Mode 2811 salted hash files (MyBB md5(md5($salt).md5($pass)), 5-char salts) 374MB
mdxfind-benchmark-400.zip Mode 400 PHPBB3/phpass hash files (100K hashes, variable iterations) 2.3MB
mdxfind-benchmark-1500.zip Mode 1500 DESCRYPT hash files (1M hashes, 4096 salts) 15MB

Test files

The benchmark hash files were generated from rockyou.txt using mdxfind itself:

mdxfind -z -h '^MD5$' -f /dev/null rockyou.txt | cut -d' ' -f2 | cut -d: -f1 > testfull.txt

Unsalted MD5 test files

File Hashes Solvable Description
testfull.txt 14,341,564 100% All hashes are solvable with rockyou.txt
test50.txt 14,341,564 ~50% 50% of hashes randomly reversed (unsolvable noise)
test10.txt 14,341,564 ~10% 90% of hashes randomly reversed (mostly noise)
sm-testfull.txt 1,000,000 100% First 1M lines of testfull.txt (for small/ARM hosts)
sm-test50.txt 1,000,000 ~50% First 1M lines of test50.txt
sm-test10.txt 1,000,000 ~10% First 1M lines of test10.txt

Salted MD5 test files

Each line is hash:salt format, where the hash is MD5(original_MD5_hash + salt) and the salt is a random 3-character string. This creates a double-hash-plus-salt scheme: solving requires computing MD5(candidate), appending the salt, then computing MD5 again — effectively MD5(MD5($pass) + $salt). This is internal type e31 (MD5SALT with MD5 pre-hash).

This creates two classes of solvable hashes within each salted file:

The standard salted benchmark uses e31 (MD5SALT) only, for fair comparison with other tools:

mdxfind -M e31 -F saltfull.txt rockyou.txt

However, mdxfind can solve both types in a single run by adding e541:

mdxfind -M e31,e541 -F saltfull.txt rockyou.txt

To our knowledge, no other tool supports the reversed-MD5-salted variant (e541).

File Hashes Solvable Description
saltfull.txt 14,341,564 100% Salted version of testfull.txt
salt50.txt 14,341,564 ~50% Salted version of test50.txt
salt10.txt 14,341,564 ~10% Salted version of test10.txt
sm-saltfull.txt 1,000,000 100% First 1M lines (for small/ARM hosts)
sm-salt50.txt 1,000,000 ~50% First 1M lines
sm-salt10.txt 1,000,000 ~10% First 1M lines

The reversed hashes simulate real-world conditions where only a fraction of the hash list is solvable with a given wordlist. The test10.txt scenario (10% solvable) is typical of working with large leaked hash collections.

Mode 2811 (MyBB) salted test files

Hashcat mode 2811 / mdxfind type e367 (MD5-MD5SALTMD5PASS): md5(md5($salt).md5($pass)) with 5-character random printable ASCII salts. Generated from rockyou.txt using gen2811 (source in bench/).

Note: The 2811.zip was updated 2026-04-05 — the previous version was missing the first salted password from rockyou.txt.

File Hashes Solvable Description
salt2811.txt 14,341,564 100% Full rockyou set, compound salted
sm-salt2811.txt 1,000,000 100% First 1M lines (for small/ARM hosts) 
# Full 2811 benchmark
time mdxfind -M 2811 -F salt2811.txt rockyou.txt > /dev/null

# Small 2811 benchmark
time mdxfind -M 2811 -F sm-salt2811.txt rockyou.txt > /dev/null

Running the benchmark

# Full unsalted benchmark
gunzip rockyou.txt.gz
time mdxfind -f testfull.txt rockyou.txt > /dev/null
time mdxfind -f test50.txt rockyou.txt > /dev/null
time mdxfind -f test10.txt rockyou.txt > /dev/null

# Full salted benchmark
time mdxfind -M e31 -F saltfull.txt rockyou.txt > /dev/null
time mdxfind -M e31 -F salt50.txt rockyou.txt > /dev/null
time mdxfind -M e31 -F salt10.txt rockyou.txt > /dev/null

# Small unsalted benchmark (for Raspberry Pi, etc.)
time mdxfind -f sm-testfull.txt rockyou.txt > /dev/null
time mdxfind -f sm-test50.txt rockyou.txt > /dev/null
time mdxfind -f sm-test10.txt rockyou.txt > /dev/null

# Small salted benchmark
time mdxfind -M e31 -F sm-saltfull.txt rockyou.txt > /dev/null
time mdxfind -M e31 -F sm-salt50.txt rockyou.txt > /dev/null
time mdxfind -M e31 -F sm-salt10.txt rockyou.txt > /dev/null

Report: CPU model, OS, thread count, wall-clock time, and hashes found for each test file.

Standard Benchmark Results

Unsalted MD5 (rockyou.txt wordlist, sorted by testfull time)

Full test (14.3M hashes)

Expected finds: testfull=14,341,564, test50=7,169,180, test10=1,434,116.

Machine CPU Clock Full 50% 10% Rate (full)
dev3 Apple M2 Max (12 cores) 3.5 GHz 3.0s 2.0s 1.0s 4.7M/s
dev1 Apple M1 (8 cores) 3.2 GHz 3.0s 2.0s 1.0s 4.7M/s
mmt 2x Xeon E5-2697 v4 (72T) 2.3 GHz 8.0s 4.0s 1.0s 1.8M/s
firefly AArch64 RK3399 (6 cores) 2.0 GHz 9.0s 7.0s 4.0s 1.6M/s
ubpower8 POWER8 (8 cores) 3.4 GHz 29.0s 14.0s 3.0s 0.5M/s

Small test (1M hashes)

Expected finds: sm-testfull=1,000,000, sm-test50=500,583, sm-test10=100,203.

Machine CPU Clock Full 50% 10% Rate (full)
firefly AArch64 RK3399 (6 cores) 2.0 GHz 3.0s 3.0s 3.0s 4.8M/s
pi3 ARMv7 BCM2837 (4 cores) 1.2 GHz 7.0s 6.0s 6.0s 2.0M/s
pi1a ARMv6 BCM2835 (1 core) 700 MHz 87.0s 76.0s 69.0s 0.16M/s

Salted MD5SALT (e31) — sm-saltfull (1M hashes, 1M unique salts, rockyou.txt wordlist)

Machine CPU/GPU Clock Found Time Hash calcs Rate
-- 12x NVIDIA RTX 4090 OpenCL (64 cores) -- 1,000,000 14s 1,981B 328G/s
ioblade 5-GPU OpenCL (2x AMD gfx1201 + RTX 4070 Ti + RTX 3080 + AMD iGPU) -- 1,000,000 31s 1,831B 62.9G/s
mmt NVIDIA RTX 4070 Ti SUPER OpenCL + CPU -- 1,000,000 45s 353B 8.03G/s
dev3 Apple M2 Max Metal (12 cores) 3.5 GHz 1,000,000 262s 697B 2.65G/s
fpga NVIDIA GTX 1080 OpenCL -- 1,000,000 376s 353B 940M/s
gp1 AMD Radeon HD 7950 OpenCL + CPU -- 1,000,000 603s 353B 586M/s
mmt 2x Xeon E5-2697 v4 (72T) 2.3 GHz 1,000,000 686s 351B 513M/s
hpi7 NVIDIA GTX 960 OpenCL -- 1,000,000 902s 353B 391M/s
-- 12x NVIDIA RTX 4090 hashcat 7.1 (Pure Kernel) -- 1,000,000 996s 8,461B 14.3G/s
dev1 Apple M1 Metal (8 cores) 3.2 GHz 1,000,000 1035s 695B 672M/s
gp NVIDIA Tesla M2070 OpenCL + CPU -- 1,000,000 1616s 353B 219M/s
mmt NVIDIA RTX 4070 Ti SUPER hashcat 6.2.6 (Pure Kernel) -- 1,000,000 1661s -- 1.53G/s
dev3 Apple M2 Max CPU (12 cores) 3.5 GHz 1,000,000 2635s 351B 133M/s
fpga NVIDIA GTX 1080 hashcat (Pure Kernel) -- 1,000,000 4404s -- 175.9M/s
dev1 Apple M1 CPU (8 cores) 3.2 GHz 1,000,000 4859s 351B 72.3M/s
hpi7 NVIDIA GTX 960 hashcat 6.2.6 (Pure Kernel) -- 1,000,000 8191s -- 60.8M/s
ubpower8 POWER8 (80T) 3.4 GHz 1,000,000 12483s 961B 77.0M/s

Salted MD5-MD5SALTMD5PASS (e367, mode 2811) — sm-salt2811 (1M hashes, 1M unique salts, rockyou.txt wordlist)

Algorithm: md5(md5($salt).md5($pass)) — compound salted type requiring three MD5 operations per candidate per salt.

mdxfind -M 2811 -F bench/sm-salt2811.txt rockyou.txt > /dev/null
hashcat -a 0 -m 2811 -o /dev/null --potfile-disable sm-salt2811.txt rockyou.txt
Machine CPU/GPU Clock Found Time Hash calcs Rate
mmt NVIDIA RTX 4070 Ti SUPER OpenCL + CPU -- 1,000,000 83s 503B 6.25G/s
dev3 Apple M2 Max Metal + CPU 3.5 GHz 1,000,000 191s 503B 2.65G/s
fpga NVIDIA GTX 1080 OpenCL + CPU -- 1,000,000 447s 503B 1.13G/s
gp1 AMD Radeon HD 7950 OpenCL + CPU -- 1,000,000 458s 503B 1.10G/s
dev1 Apple M1 Metal + CPU 3.2 GHz 1,000,000 757s 503B 665M/s
hpi7 NVIDIA GTX 960 OpenCL + CPU -- 1,000,000 1140s 503B 441M/s
mmt NVIDIA RTX 4070 Ti SUPER hashcat 6.2.6 (Pure Kernel) -- 1,000,000 2131s -- 2.03G/s
gp NVIDIA Tesla M2070 OpenCL + CPU -- 1,000,000 2713s 503B 186M/s
fpga NVIDIA GTX 1080 hashcat 6.2.6 (Pure Kernel) -- 1,000,000 7740s -- 169.1M/s
hpi7 NVIDIA GTX 960 hashcat 6.2.6 (Pure Kernel) -- 1,000,000 10020s -- 76.5M/s

PHPBB3/phpass (e455, mode 400) — sm-salt400 (100K hashes, 100K passwords)

Algorithm: iterated MD5 with per-hash salt and variable iteration count (512-2048 rounds). Due to the high per-hash computational cost, this benchmark uses 100,000 hashes and the first 100,000 passwords from rockyou.txt rather than the full 14.3M wordlist.

mdxfind -M 400 -F bench/sm-salt400.txt bench/sm-salt400.pass > /dev/null
hashcat -a 0 -m 400 -o /dev/null --potfile-disable sm-salt400.txt sm-salt400.pass
Machine CPU/GPU Clock Found Time Hash calcs Rate
mmt NVIDIA RTX 4070 Ti SUPER OpenCL + CPU -- 100,000 64s 2,678B 42.5G/s
fpga NVIDIA GTX 1080 OpenCL + CPU -- 100,000 180s 2,676B 14.9G/s
dev3 Apple M2 Max Metal + CPU 3.5 GHz 100,000 304s 2,795B 9.20G/s
mmt NVIDIA RTX 4070 Ti SUPER hashcat 6.2.6 (Pure Kernel) -- 100,000 324s -- 19.4M/s
fpga NVIDIA GTX 1080 hashcat 6.2.6 (Pure Kernel) -- 100,000 578s -- 5.98M/s
hpi7 NVIDIA GTX 960 OpenCL + CPU -- 100,000 738s 2,675B 3.63G/s
dev1 Apple M1 Metal + CPU 3.2 GHz 100,000 1196s 2,792B 2.33G/s
hpi7 NVIDIA GTX 960 hashcat 6.2.6 (Pure Kernel) -- 100,000 1863s -- 2.31M/s

DESCRYPT (e500, mode 1500) — sm-salt1500 (1M hashes, 4096 salts, 1M passwords)

Algorithm: traditional DES crypt() with 12-bit salt and 8-character maximum password length.

mdxfind -M 1500 -F bench/sm-salt1500.txt bench/sm-salt1500.pass > /dev/null
hashcat -a 0 -m 1500 -o /dev/null --potfile-disable sm-salt1500.txt sm-salt1500.pass
Machine CPU/GPU Clock Found Time Hash calcs Rate
mmt NVIDIA RTX 4070 Ti SUPER OpenCL + CPU -- 1,000,000 9s 4B 453M/s
fpga NVIDIA GTX 1080 OpenCL + CPU -- 1,000,000 26s 4B 170M/s
dev3 Apple M2 Max Metal + CPU 3.5 GHz 1,000,000 54s 4B 75.9M/s
fpga NVIDIA GTX 1080 hashcat 6.2.6 (Pure Kernel) -- 1,000,000 84s -- 34.0M/s
hpi7 NVIDIA GTX 960 OpenCL + CPU -- 1,000,000 103s 4B 40.6M/s
mmt NVIDIA RTX 4070 Ti SUPER hashcat 6.2.6 (Pure Kernel) -- 1,000,000 135s -- 126.6M/s
dev1 Apple M1 Metal + CPU 3.2 GHz 1,000,000 141s 4B 29.0M/s
hpi7 NVIDIA GTX 960 hashcat 6.2.6 (Pure Kernel) -- 1,000,000 221s -- 4.36M/s
fpga AMD Ryzen 7 1800X CPU (16T) 3.6 GHz 1,000,000 1378s 4B 2.97M/s
dev3 Apple M2 Max CPU (12 cores) 3.5 GHz 1,000,000 1849s 4B 2.21M/s
hpi7 Intel Xeon E3-1230 v5 CPU (8T) 3.4 GHz 1,000,000 2390s 4B 1.71M/s

MD5 with mask (e1, mode 0) — testfull (14.3M hashes, rockyou.txt + ?d?d mask)

Algorithm: plain MD5 with a 2-digit mask appended to each candidate, expanding the keyspace by 100x. This tests hybrid mask attack performance with a large hash set.

mdxfind -M e1 -F testfull.txt -n '?d?d' rockyou.txt > /dev/null
hashcat --potfile-disable -a 6 -m 0 -o /dev/null testfull.txt rockyou.txt '?d?d'
Machine CPU/GPU Clock Found Time Hash calcs Rate
mmt NVIDIA RTX 4070 Ti SUPER OpenCL + CPU -- 2,933,365 6s 1.4B 413M/s
fpga NVIDIA GTX 1080 OpenCL + CPU -- 2,933,365 7s 1.4B 286M/s
hpi7 NVIDIA GTX 960 OpenCL + CPU -- 2,933,365 8s 1.4B 239M/s
mmt 2x Xeon E5-2697 v4 (72T) 2.3 GHz 2,933,365 10s 1.4B 159M/s
dev3 Apple M2 Max Metal + CPU 3.5 GHz 2,933,365 61s 1.4B 24.4M/s
dev1 Apple M1 Metal + CPU 3.2 GHz 2,933,365 67s 1.4B 21.5M/s
hpi7 NVIDIA GTX 960 hashcat 6.2.6 (Pure Kernel) -- 2,933,365 115s -- 12.4M/s
fpga NVIDIA GTX 1080 hashcat 6.2.6 (Pure Kernel) -- 2,933,365 158s -- 9.10M/s
dev3 Apple M2 Max CPU (12 cores) 3.5 GHz 2,933,365 163s 1.4B 8.87M/s
dev1 Apple M1 CPU (8 cores) 3.2 GHz 2,933,365 182s 1.4B 7.93M/s
fpga AMD Ryzen 7 1800X CPU (16T) 3.6 GHz 2,933,365 222s 1.4B 6.49M/s
mmt NVIDIA RTX 4070 Ti SUPER hashcat 6.2.6 (Pure Kernel) -- 2,933,365 265s -- 5.60M/s

The salted benchmark is dramatically more expensive than unsalted because each candidate must be tested against every unique salt. With 1M unique salts and 14.3M passwords, this requires hundreds of billions of hash computations.

Notes: - The M1 processes the full 14.3M hash set in 3 seconds — hash loading time dominates. - Lower solvability (test10) runs faster because fewer hash matches trigger output processing. - The 72-core Xeon's per-thread rate is modest, but thread count gives it strong absolute throughput on salted workloads. - The ARMv6 Pi 1 is ~30x slower than the M1 but still functional for smaller hash sets.

Community Benchmarks

mdxfind vs hashcat vs john — 25M hashes, 10 hash types

Contributed by @A1131. 25,000,000 hashes, 200MB wordlist, Ubuntu 24.04. mdxfind tested all 10 types (MD5 through SHA256) simultaneously via -m e1-e10.

time ./mdxfind -m e1-e10 -f hashes.txt wordlist.txt
Tool Hardware Time
mdxfind Intel Core i5-9300H (CPU) 26.0s
john RTX 1050 Ti (GPU) 31.1s
hashcat RTX 1050 Ti (GPU) 56.9s

mdxfind on a laptop CPU outperformed both GPU-accelerated tools on a mid-range GPU — while simultaneously testing 10 hash types. This reflects mdxfind's architecture: it loads all hashes into a Judy array and tests every candidate against the entire hash set in a single pass, whereas hashcat and john are optimized for smaller hash lists with deeper iteration counts.

mdxfind's advantage grows with hash list size — the Judy array lookup is O(1) regardless of whether there are 1,000 or 100,000,000 hashes loaded.

Adding Your Benchmarks

If you have benchmark results comparing mdxfind to other tools, please open an issue or pull request. Include: