I used it as a base to run Stable Diffusion on an AMD APU (a Ryzen 6900HX, with integrated Radeon 680M, accompanied by 32GB of DDR5 RAM).
Except I used archlinux, pytorch version: 2.4.1+rocm6.1 and using webui's "forge" fork.
Kernel 6.10 seems to have brought some magic for APU memory sharing, and it seems better to rely on it than on a definition of UMA allocation in the BIOS.
I noticed that if I set UMA_SPECIFIED to 16GB in my BIOS (the option is available), I ended up with 16GB of VRAM recognized by the OS, but a limit of about 8GB of GTT memory, and it's this memory that seems to be consider as VRAM for Stable-Diffusion/webui.
By defining UMA_AUTO, I got 512MB of VRAM according to dmesg, but 16GB of GTT memory, and then Stable-Diffusion/forge-webui considers those 16GB as usable VRAM.
(with UMA_GAME_OPTIMIZED I got about 13~14GB of usable GTT memory)
I use UMA_AUTO to maximize the available memory and was able to generate both SDXL images in your article in 4m40s and 3m31s respectively.
In Europe, the price of the machine is currently around 439€ (in the 16GB DDR5 RAM version), and you need to add around 100€ to buy 2x16GB of replacement memory... I think it's a bit cheaper in the US, and models with 32GB should be available.
There is a price drop, due to new-gen APU AI-oriented (probably not linux-optimized, and 2~3x more expensive)
One question can we use the cheep used Tesla cards(No graphics out) . They go for 100-150$ on eBay with 16GB VRAM. Reason I am asking is it looks quite attractive and good alternative to 2000$ graphics cards. I agree they are not the latest and greatest but won’t they add a decent alternative to built-in graphics.
One other issue I have is using RAM as VRAM is the CPU is throttled with GPU because both are using same RAM bus.
Thanks a lot, great article! Do you think it is going to work on more power efficient 5700U with comparable efficiency? I’m planning to run it through vm alongside other projects like Home Assistant. Will it be ok?
Comments about How to Build a $300 AI Computer for the GPU-Poor
I used it as a base to run Stable Diffusion on an AMD APU (a Ryzen 6900HX, with integrated Radeon 680M, accompanied by 32GB of DDR5 RAM).
Except I used archlinux, pytorch version: 2.4.1+rocm6.1 and using webui's "forge" fork.
Kernel 6.10 seems to have brought some magic for APU memory sharing, and it seems better to rely on it than on a definition of UMA allocation in the BIOS.
I noticed that if I set UMA_SPECIFIED to 16GB in my BIOS (the option is available), I ended up with 16GB of VRAM recognized by the OS, but a limit of about 8GB of GTT memory, and it's this memory that seems to be consider as VRAM for Stable-Diffusion/webui.
By defining UMA_AUTO, I got 512MB of VRAM according to dmesg, but 16GB of GTT memory, and then Stable-Diffusion/forge-webui considers those 16GB as usable VRAM.
(with UMA_GAME_OPTIMIZED I got about 13~14GB of usable GTT memory)
I use UMA_AUTO to maximize the available memory and was able to generate both SDXL images in your article in 4m40s and 3m31s respectively.
In Europe, the price of the machine is currently around 439€ (in the 16GB DDR5 RAM version), and you need to add around 100€ to buy 2x16GB of replacement memory... I think it's a bit cheaper in the US, and models with 32GB should be available.
There is a price drop, due to new-gen APU AI-oriented (probably not linux-optimized, and 2~3x more expensive)
😊+
Hay, I am new to AI workstations.
One question can we use the cheep used Tesla cards(No graphics out) . They go for 100-150$ on eBay with 16GB VRAM. Reason I am asking is it looks quite attractive and good alternative to 2000$ graphics cards. I agree they are not the latest and greatest but won’t they add a decent alternative to built-in graphics.
One other issue I have is using RAM as VRAM is the CPU is throttled with GPU because both are using same RAM bus.
Thanks for the starting setup help
GP
😊+
Thanks a lot, great article! Do you think it is going to work on more power efficient 5700U with comparable efficiency? I’m planning to run it through vm alongside other projects like Home Assistant. Will it be ok?
😊+