hashcat Forum
Tenstorrent Grayskull and Wormhole - better or worse than a GPU? - Printable Version

+- hashcat Forum (https://hashcat.net/forum)
+-- Forum: Misc (https://hashcat.net/forum/forum-15.html)
+--- Forum: Hardware (https://hashcat.net/forum/forum-13.html)
+--- Thread: Tenstorrent Grayskull and Wormhole - better or worse than a GPU? (/thread-12097.html)



Tenstorrent Grayskull and Wormhole - better or worse than a GPU? - Have Blue - 07-29-2024

Just came across these boards designed for AI models that use a whole bunch of RISC-V cores:
https://tenstorrent.com/hardware/grayskull
https://tenstorrent.com/hardware/wormhole

There doesn't appear to be an OpenCL package available for the processors, just their own high- and low-level APIs.  Performance for the Grayskull e150 is stated as 332 TeraFLOPS, while a 4090 should in theory do 1.3 PetaFLOPS with the same FP8 benchmark.  At $800 for an e150 vs $1800 for a 4090, the 4090 is easily the most performance per dollar for FP8.

Still, I'm intrigued by alternate hardware possibilities for hashcat - FPGAs, ASICs, etc.  Could a Tenstorrent card provide any sort of advantage were hashcat able to work with one?


RE: Tenstorrent Grayskull and Wormhole - better or worse than a GPU? - Chick3nman - 08-06-2024

It's possible that you could get some algorithms working, I've explored alternative processors quite a few times. That being said, you've already identified the reason why we typically don't spend much time on it. The 4090(and really GPUs in general) remains the best performance per dollar, even at it's relatively high MSRP. Beyond that, it's well supported and easy to code for, and the code is portable and not device specific. FPGAs are becoming worthwhile in a few cases, and you will see support for those as hashcat grows, but those cases are extremely specific(bcrypt, scrypt, etc.) and FPGA devices are both very expensive and also fairly difficult to write kernels for. My guess is that a "many-core" chip is probably one of the most decent alternative processor to play with, but I don't expect it to perform very well outside of maybe a few edge cases with anti-GPU algorithms.