Hacker News new | past | comments | ask | show | jobs | submit login

>Fortunately the slowdown of minGPT training is only ~3% with this setup.

The cards are run at a very unfavorable part of Freq/Voltage curve. Increasing freq. (performance) effectively scales the power in a cubic manner.




To be pedantic 1.03 * 3 is ~1.093, which is much less than 1.4 = 350/250, so the behavior is still quite surprising.




Join us for AI Startup School this June 16-17 in San Francisco!

Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: