3 ms·

Does this run on AMD? A problem I see is that a lot of times everything works fine on rocm+hip, but since nvidia dominates the machine learning market (and thu

Does this run on AMD? A problem I see is that a lot of times everything works fine on rocm+hip, but since nvidia dominates the machine learning market (and thus most researches run nvidia), most forks don't bother checking and just advertise compatibility with nvidia and sometimes apple M1. Problem is, AMD GPUs are much cheaper!

cypress66 4y ago
> Problem is, AMD GPUs are much cheaper! Are they? I believe Nvidia (consumer) gpus have better price/performance than amd for AI.
- nextaccountic 4y ago
  I don't know about AI performance (does this happen only because of the overhead of providing CUDA through rocm+HIP?), but I was just checking and at least in my country (Brazil), for any given memory size (12GB, 8GB, 4GB) I can find cheaper AMD GPUs than NVidia GPUs Here I'm considering that the main constraint is VRAM and while stable diffusion now runs even on GPUs with 2GB RAM, there's always new developments that require more VRAM (for example, Dreambooth requires 12GB as of today)
- mrtranscendence 4y ago
  Maybe for AI? For other tasks, especially gaming, they punch well above their weight relative to Nvidia (though they lack features in comparison). It's also possible to get a 16GB card for much cheaper from AMD than Nvidia.
mrtksn 4y ago
Well, high-end stuff is always on Nvidia and Apple Silicon seems to get some love because of its unified memory that makes it possible in first place plus its popularity among developers. AMD seems to be popular among gamers on budget and the budget cards often don't have the VRAM required by default. So, AMD seems to be in this weird place where the people who can make it work don't care.
- mrtranscendence 4y ago
  For what it's worth, at the consumer level AMD cards -- at least recently -- have tended to have more VRAM than Nvidia cards. My 3080 Ti, which I bought for $1400 (though it now goes for ~$1k), has less RAM (12GB) than a 6800 XT that you can get for $600 (16GB).

cypress66 4y ago

> Problem is, AMD GPUs are much cheaper! Are they? I believe Nvidia (consumer) gpus have better price/performance than amd for AI.

nextaccountic 4y ago
I don't know about AI performance (does this happen only because of the overhead of providing CUDA through rocm+HIP?), but I was just checking and at least in my country (Brazil), for any given memory size (12GB, 8GB, 4GB) I can find cheaper AMD GPUs than NVidia GPUs Here I'm considering that the main constraint is VRAM and while stable diffusion now runs even on GPUs with 2GB RAM, there's always new developments that require more VRAM (for example, Dreambooth requires 12GB as of today)
mrtranscendence 4y ago
Maybe for AI? For other tasks, especially gaming, they punch well above their weight relative to Nvidia (though they lack features in comparison). It's also possible to get a 16GB card for much cheaper from AMD than Nvidia.

mrtksn 4y ago

Well, high-end stuff is always on Nvidia and Apple Silicon seems to get some love because of its unified memory that makes it possible in first place plus its popularity among developers. AMD seems to be popular among gamers on budget and the budget cards often don't have the VRAM required by default. So, AMD seems to be in this weird place where the people who can make it work don't care.

mrtranscendence 4y ago
For what it's worth, at the consumer level AMD cards -- at least recently -- have tended to have more VRAM than Nvidia cards. My 3080 Ti, which I bought for $1400 (though it now goes for ~$1k), has less RAM (12GB) than a 6800 XT that you can get for $600 (16GB).