Expensive AI GPUs flunk password cracking test
Hardware tests reported that $30K AI GPUs like Nvidia H200 and AMD MI300X underperformed an RTX 5090 in password‑cracking benchmarks, exposing efficiency gaps between expensive AI accelerators and consumer cards on non‑LLM workloads. The result reminds builders that hardware choice matters and that expensive AI silicon isn't uniformly superior across every task. (x.com)
A password cracker does one very simple job millions or billions of times: it guesses a password, runs the same math the login system used, and checks whether the output matches the stored hash. Hashcat is one of the best-known tools for measuring how fast a graphics processor can do that loop. (hashcat.net) That is why this result surprised people. Specops tested Nvidia’s H200, Advanced Micro Devices’ Instinct MI300X, and Nvidia’s GeForce RTX 5090, and the much cheaper RTX 5090 finished first in every benchmark they published. (specopssoft.com) On the fast, older hash called Message Digest 5, or MD5, the RTX 5090 reached 219.5 billion guesses per second. The H200 managed 124.4 billion, and the MI300X reached 164.1 billion. (specopssoft.com) On New Technology Local Area Network Manager, or NTLM, which still shows up in Windows environments, the RTX 5090 hit 340.1 billion guesses per second. The H200 posted 218.2 billion, and the MI300X posted 268.5 billion. (specopssoft.com) Even on slower, more defensive hashes, the pattern mostly held. The RTX 5090 led bcrypt at 304.8 thousand guesses per second, while the H200 reached 275 thousand and the MI300X fell to 142.3 thousand. (specopssoft.com) The expensive cards were built for a different kind of work. Nvidia says the H200 is for generative artificial intelligence and high-performance computing, and Advanced Micro Devices describes the MI300X as an accelerator for artificial intelligence, high-performance computing, and other demanding workloads. (nvidia.com) (instinct.docs.amd.com) That difference matters because password cracking is not the same as training a large language model. One job is like hauling huge pallets through a warehouse, while the other is like checking billions of tiny keys against one lock. (specopssoft.com) (nvidia.com) The H200 leans hard into giant memory pools, with 141 gigabytes of High Bandwidth Memory 3E, or HBM3E, and 4.8 terabytes per second of memory bandwidth. The RTX 5090, by contrast, ships with 32 gigabytes of Graphics Double Data Rate 7, or GDDR7, memory and starts at $1,999. (nvidia.com 1) (nvidia.com 2) The MI300X pushes that memory-heavy design even further with 192 gigabytes of High Bandwidth Memory 3, or HBM3, per accelerator. That is great when a model is too large to fit on a smaller card, but Specops’ numbers show that more memory does not automatically mean more password guesses per second. (instinct.docs.amd.com) (specopssoft.com) So the lesson is not that datacenter hardware is “bad.” The lesson is that a chip designed for artificial intelligence training can lose badly at a job that rewards a different balance of clocks, memory behavior, software support, and raw per-dollar throughput. (specopssoft.com) That is also why security teams do not look at a $30,000 accelerator and assume attackers just got ten times stronger. In this corner of computing, the benchmark says the sharper tool was the gaming card, not the one built for the artificial intelligence boom. (tomshardware.com) (specopssoft.com)
