How to Run Deepseek-R1-0528 Locally (unsloth.ai)
from yogthos@lemmy.ml to technology@lemmy.ml on 29 May 00:38
https://lemmy.ml/post/30822277

#technology

threaded - newest

N0x0n@lemmy.ml on 29 May 04:02 collapse

have at least 160GB of combined VRAM and system RAM.

Yeaaaah so it’s possible but at the same time not accessible ? Those things are fucking power hungry and now I get why we all needed to buy led things, 12 volts powered systems and switch off every standby device: AI !

And I dummy though innocently it was for the planet… 🥲

simontherockjohnson@lemmy.ml on 29 May 05:06 next collapse

Lol the suggested hardware for usable performance is $50k for the GPUs alone and that’s an SXM5 socket so all proprietary extremely expensive and specific hardware.

My PC currently has a 7900 XTX which gives me about 156 GB combined VRAM, but it literally generates 1-3 words per second even at this level. DDR5 wouldn’t really help, because it’s a memory bandwidth issue.

TBH for most reasonable use cases 8 bit parameter size quantizations that can run on a laptop will give you more or less what you want.

corsicanguppy@lemmy.ca on 29 May 06:59 collapse

And I [thought?] it was for the planet… 🥲

It was maybe never for the planet. We accidentally got LED lights, but the entire effort may have been to provide a scapegoat for the energy issues the oil barons didnt want to address. Dig a bit and you’ll find some neat charts and timelines, and maybe you’ll wonder too whether it was all a smokescreen like the recycling thing sometimes resembles.