CoreML Stable Diffusion | Hacker News Reader

CoreML Stable Diffusion

github.com

52 pointsrcarmo5 comments

It's strange to me how much effort has gone into porting things over to run on Apple Silicon while AMD GPUs seem to have been completely abandoned as an ML target. It's understandable that Apple would promote their own CPUs of course, but other than Metal they don't seem to have any interest in the GPU side, and few ML tools use Metal as a backend since Nvidia has had a dominant position in the GPU space.

AMD dropped the ball real bad

I've been running SD on AMD hardware since day one. It's a little bit of a hassle compared to NVIDIA for sure, but I attribute that to a decade or two of somewhat shady (and very successful) moves to make ML researchers need/want to use NVIDIA.

AMD is behind ROCm, and, development is active. So, maybe you can clarify what you mean by "abandoned"? I'm not feeling it.

https://en.m.wikipedia.org/wiki/ROCm

Any benchmarks out there preparing Nvidia versus AMD?

ROCm's great for data centers, but good luck finding anything about desktop GPUs on their site apart from this lone blog post: https://community.amd.com/t5/instinct-accelerators/exploring...

There's a good explanation of AMD's ROCm targets here: https://news.ycombinator.com/item?id=28200477

It's currently a PITA to get common Python libs like Numba to even talk to AMD cards (admittedly Numba won't talk to older Nvidia cards either and they deprecate ruthlessly; I had to downgrade 8 versions to get it working with a 5yo mobile workstation). YC-backed Ivy claims to be working on unifying ML frameworks in a hardware-agnostic way but I don't have enough experience to assess how well they're succeeding yet: https://lets-unify.ai

I was happy to see DiffusionBee does talk the GPU in my late-model intel Mac, though for some reason it only uses 50% of its power right now. I'm sure the situation will improve as Metal 3.0 and Vulkan get more established.

Reply on news.ycombinator.com