Conversation

Notices

1. Embed this notice
   Sick Sun (sun@shitposter.world)'s status on Thursday, 01-Aug-2024 03:36:11 JST Sick Sun
   I think I am going to reinstall the entire OS on this machine because the proprietary nvidia drivers keep fucking up
   • Embed this notice
     Sick Sun (sun@shitposter.world)'s status on Thursday, 01-Aug-2024 04:03:49 JST Sick Sun

     in reply to

     • Bricky
     @thatbrickster not for AI GPU stuff
   • Embed this notice
     Bricky (thatbrickster@shitposter.world)'s status on Thursday, 01-Aug-2024 04:03:50 JST Bricky

     in reply to
     @sun I hear the open drivers have reached performance parity, assuming Linux.
   • Embed this notice
     Sick Sun (sun@shitposter.world)'s status on Thursday, 01-Aug-2024 04:05:16 JST Sick Sun

     in reply to

     • बुद्धीविनाषक् भेड़ मीपर् [Meeper]
     @meeper I'm doing AI stuff using nvidia CUDA drivers
   • Embed this notice
     बुद्धीविनाषक् भेड़ मीपर् [Meeper] (meeper@udongein.xyz)'s status on Thursday, 01-Aug-2024 04:05:17 JST बुद्धीविनाषक् भेड़ मीपर् [Meeper]

     in reply to
     @sun wasn't novideo finally open sourcing and fixing their shit?
   • Embed this notice
     Bricky (thatbrickster@shitposter.world)'s status on Thursday, 01-Aug-2024 04:09:19 JST Bricky

     in reply to
     @sun Ah, fair enough. I've become re-interested but unsure where to start. Many tools, little on DIY.
   • Embed this notice
     Sick Sun (sun@shitposter.world)'s status on Thursday, 01-Aug-2024 04:09:19 JST Sick Sun

     in reply to

     • Bricky
     @thatbrickster if you have a discrete nvidia gpu I think you can find models that will run on how much RAM you have, not sure. otherwise if you want to experiment you can run smaller models on a CPU and it's just kind of slow.
   • Embed this notice
     Sick Sun (sun@shitposter.world)'s status on Thursday, 01-Aug-2024 04:11:49 JST Sick Sun

     in reply to

     • बुद्धीविनाषक् भेड़ मीपर् [Meeper]
     • dirb
     @dirb @meeper Oh cool, I'll find out if the software supports it, I'd rather not use the nvidia software if possible because it's really shitty
   • Embed this notice
     dirb (dirb@pleroma.dirb.xyz)'s status on Thursday, 01-Aug-2024 04:11:51 JST dirb

     in reply to

     • बुद्धीविनाषक् भेड़ मीपर् [Meeper]
     @sun @meeper according to phoronix the open source kernel modules should be at parity for cuda stuff too
     https://www.phoronix.com/review/nvidia-555-open
   • Embed this notice
     mangeurdenuage :gnu: :trisquel: :gondola_head: 🌿 :abeshinzo: :ignucius: (mangeurdenuage@shitposter.world)'s status on Thursday, 01-Aug-2024 04:58:36 JST mangeurdenuage :gnu: :trisquel: :gondola_head: 🌿 :abeshinzo: :ignucius:

     in reply to
     @sun I recall reading the recent "open source" drivers nvidia distributes now are equal to the proprietary one.
     Sick Sun likes this.
   • Embed this notice
     Sick Sun (sun@shitposter.world)'s status on Thursday, 01-Aug-2024 05:13:47 JST Sick Sun

     in reply to

     • Bricky
     @thatbrickster ohhh yeah I am learning how to do that. one hint I can give you is training a model from scratch is probably a lot less feasible than finding one and fine-tuning it.
   • Embed this notice
     Bricky (thatbrickster@shitposter.world)'s status on Thursday, 01-Aug-2024 05:13:48 JST Bricky

     in reply to
     @sun I have a machine with a discrete GPU but that wasn't quite the question. I'm aware of the models, I'm not aware of how to train my own from scratch.
   • Embed this notice
     翠星石 (suiseiseki@freesoftwareextremist.com)'s status on Thursday, 01-Aug-2024 20:20:51 JST 翠星石

     in reply to

     • बुद्धीविनाषक् भेड़ मीपर् [Meeper]
     • Bricky
     @meeper @sun @thatbrickster There is no "open drivers", just a MIT expat shim that facilitates GPLv2-infringement of the driver that runs on the GPU.
     
     Previously there was a GPLv2-infringing shim and part of the driver ran in userspace and another part ran on the GPU - so there is no real difference.
     
     
     There is a free software driver available called nouveau, although any re-clocking support is via reverse engineering and only up to the 700 series has free peripheral software for the GPUs - anything after the 700 series is designed to maliciously refuse to spin up the fans unless proprietary software that passes a signature check (to make it cryptographically impossible to replace it with free software) is loaded up onto the card.