Specs:
i7 4790K ($50)
9020 Optiplex Motherboard ($10)
32GB DDR3 RAM ($35)
7900 XTX 24GB VRAM ($900)
1TB M.2 NVMe ($50)
EVGA 700 BR PSU ($50)
24 pin to 8 pin adapter for motherboard ($20)
NVME PCIe x4 Adapter ($12)
Molex to fan adapters ($7)
Power button adapter ($6)
Total: $1140
Using Arch GNU/Linux (Tried installing Debian, couldn't get the AMD drivers to work properly, so I'm using Arch for the moment)
Can run Stable Diffusion, LLMs, and basically all my games at 1440p High to Ultra settings (RDR2, GTA V, Arma 3, etc.)
GPU passhtrough doesn't work for the moment, but I can still run virtual machines. This is a Haswell motherboard, which can be 100% freed in the BIOS over time when we have enough Libreboot devs who can reverse engineer the rest of the blobs. Intel ME is also disabled (theoretically, since you can't actually fully remove it). This is a build that I did for fun, in the future, I'm planning on switching to a Z690-A motherboard with DDR5(can be flashed with Dasharo firmware). If you have any questions, feel free to ask!
Is the ddr3 not hindering you in your AI training? I mean maybe not because of that massive VRAM. Just curious!
It's actually not that bad, I can run basically everything I need with good speeds. I will need more RAM in the future for larger models, but for my use case, this works just fine. I'm considering upgrading to the MSI Z690-A DDR5 motherboard, it can be flashed with Dasharo firmware but it unfortunately has more blobs (e.g. FSP binaries)
Thanks for the info!
I'm starting to believe you don't really need beefy PCs any more, with exception ofc, but those exceptions seems to be more and more rare, especially for the normal guy in his garage :-)
My kids took over my "good" PCs (modern with a good graphic card) and I'm on oldish thinkcentres/thinkpad and well it's kind of okay actually.
I even think that if I could use my mobile phone (like with 2-3 screens, ethernet...) I bet the experience would be totally acceptable.
Your card also make me remember when the trainee at work doing AI training and having the model being too big for the 16GB card the company splurged some 5.000€ for a 24GB card. Tensorflow back in 2017 or maybe 2018.
Cheers
Thanks for the info!
I'm starting to believe you don't really need beefy PCs any more, with exception ofc, but those exceptions seems to be more and more rare, especially for the normal guy in his garage :-)
My kids took over my "good" PCs (modern with a good graphic card) and I'm on oldish thinkcentres/thinkpad and well it's kind of okay actually.
I even think that if I could use my mobile phone (like with 2-3 screens, ethernet...) I bet the experience would be totally acceptable.
Your card also make me remember when the trainee at work doing AI training and having the model being too big for the 16GB card the company splurged some 5.000€ for a 24GB card. Tensorflow back in 2017 or maybe 2018.
Cheers