Oddbean new post about login | logout @d @Rance :gentoo: i have a Quadro RTX 4000 in my desktop and it's way faster than my old 1070 at stable diffusion but i vaguely remember seeing some benchmarks of game performance where the two weren't that far apart
@09085299 @Rance :gentoo: @d I am getting this video card for LlamaGPT and for some AI artwork. Nvidia Quadro P1000 4GB GDDR5 How good will it be?
@verita84 :Debian_logo: :firefox: :bing: :android: ... @Rance :gentoo: @d @09085299 Absolutely awful, the ram size matters for AI work.
@Salastil @Rance :gentoo: @d @09085299 Will it be functional but just take a little longer?
@verita84 :Debian_logo: :firefox: :bing: :android: ... @Rance :gentoo: @d @09085299 It wont be functional, pytorch has to reserve memory off the bat and when you're memory strapped it wont work. I can't run many models using my 1060 6GB because it doesn't have enough memory on the card.
@Salastil @Rance :gentoo: @d @09085299 oh fuck me.
The Quadro series is good for parallel transcoding of video, its the entire gimmick of the series.
@Salastil @Rance :gentoo: @d @09085299 well it will help the Jellyfin server at least
how come you opted for the 4gb? was it 10 bucks or something?
@Nozahagmelons @Salastil @Rance :gentoo: @d @09085299 I paid 80 each I think.....Best I could do for the OptiPlex 3050
@Nozahagmelons @Salastil @Rance :gentoo: @d @09085299 Any Alternative to stable diffusion that I can use and self-host?
not to my knowledge, but KoboldAI is pretty cool for text generation
@Nozahagmelons @Salastil @Rance :gentoo: @d @09085299 Nice, better than Llama?
What? You're too good to just walk into a Microcenter and steal a RTX 4090ti? You'll never get ahead in Biden's America until you start playing by the new rules. https://media.salastil.com/media/b247a6e82ca21ed1fda4dd9ee783e7c1d37c5805644afc96e2f4c415e456e9ec.png
Even my old ass Quadro M4000 has 8 GB VRAM and can absolutely own any modern game on ultra settings even shit optimized games like fallout 76
@Salastil @Rance :gentoo: @d @09085299 So if you use AI with a CPU, it loads all the models into RAM from the start. When you use a GPU, are you saying it loads it all on the GPU instead of RAM?
you can pick up these Tesla K40 12GB cards I'm talking about for like $40 they're strictly for compute though, no video output
@d @Rance :gentoo: @09085299 it's a old Dell from 2016/2015 . Low-Profile slots
