
Will spare AI accelerators make password cracking dramatically cheaper? Specops Software benchmarked Nvidia's H200 and AMD's MI300X, both data-centre GPUs costing $30,000 or more, against Nvidia's consumer RTX 5090. The consumer card won on every algorithm tested. For developers, the takeaway isn't about hardware. Your password hashing choice matters far more than the attacker's GPU.
What happened
In a blog post published in late March 2026 and syndicated by BleepingComputer in April, Specops compared the three GPUs on five hash types, using Hashcat-style throughput figures:
- MD5: H200 124.4 GH/s, MI300X 164.1 GH/s, RTX 5090 219.5 GH/s
- NTLM: H200 218.2 GH/s, MI300X 268.5 GH/s, RTX 5090 340.1 GH/s
- SHA-256: H200 15.1 GH/s, MI300X 24.7 GH/s, RTX 5090 27.7 GH/s
- SHA-512: H200 5.2 GH/s, MI300X 8.8 GH/s, RTX 5090 10.0 GH/s
- bcrypt: H200 275 kH/s, MI300X 142.3 kH/s, RTX 5090 304.8 kH/s
Specops puts the result down to design. AI accelerators are built for low-precision matrix maths and very high memory bandwidth, while hash cracking mostly needs fast integer operations, which consumer GPUs handle well. Its conclusion: AI hardware is no special new threat to password hashing, but a single consumer GPU already tears through weak, fast hashes.
Specops sells password-policy products, and the post doesn't disclose the bcrypt cost factor, Hashcat version or test host. Treat the numbers as indicative, not a lab-grade benchmark.
Who is affected
- Any application that stores passwords, especially legacy systems still using MD5, SHA-1, or unsalted or single-round SHA-256/512.
- Windows and Active Directory environments, where NTLM hashes are a frequent target once domain controllers or memory dumps are compromised.
- Users with short or reused passwords, whose hashes fall quickly whatever the hardware.
What it means for developers and engineering teams
The gap between rows matters more than the gap between GPUs. A single RTX 5090 tries about 220 billion MD5 guesses per second, but only about 300 thousand bcrypt guesses, roughly six orders of magnitude fewer. A modest change in your code buys you far more than any change in attacker hardware takes away.
Fast hashes (MD5, SHA-1, SHA-256, SHA-512, NTLM) are built for speed, so they suit integrity checks, not passwords. Even salted, an 8-character password hashed with SHA-256 falls to a single card within days at most. Memory-hard algorithms like Argon2id and scrypt also blunt GPUs and accelerators, because each guess needs a large amount of RAM, so parallel throughput collapses.
The result also cuts the other way. Specops' finding that AI chips are no better than gaming cards doesn't mean you're safe. It means cheap consumer hardware, rentable by the hour, is already the threat model.
What to do now
- Use Argon2id for new systems. OWASP's minimum is 19 MiB memory, 2 iterations and parallelism 1 (
m=19456, t=2, p=1). For example, in Node:import argon2 from "argon2"; const hash = await argon2.hash(pw, { type: argon2.argon2id, memoryCost: 19456, timeCost: 2, parallelism: 1 }); - If you're on bcrypt, use a work factor of at least 10 (higher if your login latency budget allows), and enforce the 72-byte input limit.
- Migrate legacy hashes on login. Verify against the old hash, then re-hash with Argon2id. Or wrap old hashes now (for example
argon2id(md5_hash)) so the whole table is protected at once. - Allow long passphrases and block breached passwords at sign-up and password change. Specops recommends at least 15 characters.
- Add MFA, ideally phishing-resistant, so a cracked password alone isn't enough.
- In Windows estates, reduce NTLM use, protect domain controller backups, and keep privileged accounts on long, unique passwords.