[rwmj]

Stable Diffusion on a 16 core AMD CPU takes for me about 2-3 hours to generate an image, just to give you a rough idea of the performance. (On the same AMD's iGPU it takes 2 minutes or so).

[weebull]

WTF!

On my 5900X, so 12 cores, I was able to get SDXL to around 10-15 minutes. I did do a few things to get to that.

1. I used an AMD Zen optimised BLAS library. In particular the AMDBLIS one, although it wasn't that different to the Intel MKL one.

2. I preload the jemalloc library to get better aligned memory allocations.

3. I manually set the number of threads to 12.

This is the start of my ComfyUI CPU invocation script.

    export OMP_NUM_THREADS=12
    export LD_PRELOAD=/opt/aocl/4.1.0/aocc/lib_LP64/libblis-mt.so:$LD_PRELOAD
    export LD_PRELOAD=/usr/lib/libjemalloc.so:$LD_PRELOAD
    export MALLOC_CONF="oversize_threshold:1,background_thread:true,metadata_thp:auto,dirty_decay_ms: 60000,muzzy_decay_ms:60000"

Honestly, 12 threads wasn't much better than 8, and more than 12 was detrimental. I was memory bandwidth limited I think, not compute.

[OJFord]

Even older GPUs are worth using then I take it?

For example I pulled a (2GB I think, 4 tops) 6870 out of my desktop because it's a beast (in physical size, and power consumption) and I wasn't using it for gaming or anything, figured I'd be fine just with the Intel integrated graphics. But if I wanted to play around with some models locally, it'd be worth putting it back & figuring out how to use it as a secondary card?

[rwmj]

One counterintuitive advantage of the integrated GPU is it has access to system RAM (instead of using a dedicated and fixed amount of VRAM). That means I'm able to give the iGPU 16 GB of RAM. For me SD takes 8-9 GB of RAM when running. The system RAM is slower than VRAM which is the trade-off here.

[OJFord]

Yeah I did wonder about that as I typed, which is why I mentioned the low amount (by modern standards anyway) on the card. OK, thanks!

[purpleflame1257]

2GB is really low. I've been able to use A111 stable diffusion on my old gaming laptop's 1060 (6GB VRAM) and it takes a little bit less than a minute to generate an image. You would probably need to try the --lowvram flag on startup.

[mat0]

No, I don't think so. I think you would need more VRAM to start with.

[smoldesu]

SDXL Turbo is much better, albeit kinda fuzzy and distorted. I was able to get decent single-sample response times (~80-100s) from my 4 core ARM Ampere instance, good enough for a Discord bot with friends.

[emadm]

Sd turbo runs nicely on a m2 MacBook Air (as does stable lm 2!)

Much faster models will come

[antman]

Which AMD CPU/iGPU are these timings for?

[rwmj]

AMD Ryzen 9 7950X 16-Core Processor

The iGPU is gfx1036 (RDNA 2).

[adrian_b]

If that is true, then the CPU variant must be a much worse implementation of the algorithm than the GPU variant, because the true ratio of the GPU and CPU performances is many times less than that.