Why do hackers want your graphics card?
A GPU (Graphics Processing Unit) runs thousands of threads at once, so attackers use it to test billions of password guesses a second against stolen hashes offline, or hijack it to mine crypto.

GPU
A GPU (Graphics Processing Unit) is the chip on a graphics card that draws the picture on your screen. To draw a frame, its rasterizer works out which pixels each triangle covers. Then a tiny program called the pixel shader colors each pixel, all in parallel. A 1920x1080 screen alone has 2,073,600 pixels, so the chip is built to repeat one small job a huge number of times at once.
A CPU (Central Processing Unit) runs a few dozen threads at once. A GPU runs thousands, because it spends more of its transistors on processing data. That makes it useful to attackers in two ways. The first is guessing passwords. The second is hijacking. In ShadowRay, reported in March 2024, attackers sent jobs to the Ray dashboard on port 8265, which had no login by design, and mined crypto on the GPUs of AI (artificial intelligence) clusters.
What a GPU actually does with passwords
An RTX 5090 splits its work across 170 streaming multiprocessors. Each one runs warps, which are groups of 32 threads that execute one instruction together. This model is called SIMT (single instruction, multiple threads). It works best when every thread follows the same steps.
Now swap pixels for passwords. Windows stores a fingerprint of your password called the NTLM (NT LAN Manager) hash. It’s one fast MD4 (Message Digest 4) pass over the password, with no salt. It’s kept in the SAM (Security Account Manager) on a single PC, or in NTDS.dit in Active Directory. Attackers steal those hashes and guess offline with hashcat: hash a word, compare, repeat. Every guess runs identical code, so all 32 threads in a warp stay busy. In a community benchmark, one RTX 5090 tested around 300 billion NTLM guesses a second. The exact figure varies with driver and tuning. It only applies to offline guessing, not to logins against a live PC.
How slow hashes protect passwords
- Choose a slow hash. Good sites use slow, salted hashes like Argon2id or bcrypt, not one fast, unsalted pass like NTLM. Current guidance lists Argon2id first and keeps bcrypt for legacy systems.
- Repeat the expensive setup. bcrypt has a cost setting. At cost 10, it repeats its expensive key setup 2^10 times, which is 1,024 times for every single guess.
- Make every guess costly. The attacker pays that price on every guess too. The GPU’s threads still run in parallel, but each one does far more work per word.
- Watch the numbers collapse. Sixteen RTX 5090s together manage only about 139 thousand bcrypt guesses a second. One card alone does around 300 billion against NTLM.
The hardware is the same. Only the hash changed. That’s why the way a site stores passwords matters.
A GPU is not a faster CPU
It’s easy to think a GPU is simply a faster CPU. It isn’t. It’s a crowd of slower workers. A CPU handles a few dozen threads at once, each free to do something different. A GPU handles thousands, grouped into warps that run the same instruction together. Password guessing fits that crowd perfectly, because every guess runs identical steps. Long, unique passwords and slow hashes take that advantage away.
What to do
- Give your email a unique password of 15 or more characters today. A GPU makes guessing cheap, but length makes it impractical.
- Turn on two-step login for your email, so a cracked password alone isn’t enough to get in.
- If you build websites, store passwords with a slow, salted hash such as Argon2id, or bcrypt at cost 10 or higher where Argon2id and scrypt aren’t available.
- If you run Ray on GPU clusters, keep its dashboard on port
8265off the open internet, since it had no login by design.
Watch it in 60 seconds
