this post was submitted on 02 Jul 2024
536 points (99.8% liked)
TechTakes
1438 readers
46 users here now
Big brain tech dude got yet another clueless take over at HackerNews etc? Here's the place to vent. Orange site, VC foolishness, all welcome.
This is not debate club. Unless it’s amusing debate.
For actually-good tech, you want our NotAwfulTech community
founded 1 year ago
MODERATORS
you are viewing a single comment's thread
view the rest of the comments
view the rest of the comments
there’s this type of reply guy on fedi lately who does the “well actually querying LLMs only happens in bursts and training is much more efficient than you’d think and nvidia says their gpus are energy-efficient” thing whenever the topic comes up
and meanwhile a bunch of major companies have violated their climate pledges and say it’s due to AI, they’re planning power plants specifically for data centers expanded for the push into AI, and large GPUs are notoriously the part of a computer that consumes the most power and emits a ton of heat (which notoriously has to be cooled in a way that wastes and pollutes a fuckton of clean water)
but the companies don’t publish smoking gun energy usage statistics on LLMs and generative AI specifically so who can say
The last part is absolutely false. The Nvidia H100 TDP is like 700W, though ostensibly configurable. The B200 is 1000W. The AMD MI300X is 750W.
They also skimp on RAM with many SKUs so you have to buy the higher clocked ones.
They run in insane power bands just to eek out a tiny bit more performance. If they ran at like a third of their power, I bet they would be at least twice as power efficient, and power use scales over nonlinearly with voltage/clock speed.
But no, just pedal to the metal. Run the silicon as hard as it can, and screw power consumption.
Other AI companies like Cerebras are much better, running at quite sane voltages. Ironically (or perhaps smartly), the Saudis invested in them.
wow, exemplary performance