[Question] Significantly Slower Performance on Apple Metal (M1) Compared to NVIDIA GPU When Using Z-Image (Q3)

[bhdrakdemr]

I’m seeing a large performance gap when running Z-Image (Q3) on Apple Metal compared to CUDA, using the same settings.
I’m new to running image generation models locally, so this may be a configuration or expectation issue.

Hardware

• RTX 1000 Ada (6 GB, CUDA): ~2 s / iteration

• MacBook Air M1 (16 GB, Metal): ~20 s / iteration

Example (5 Iterations)

• CUDA: ~2 s × 5 steps ≈ 10 seconds total

• Metal: ~20 s × 5 steps ≈ 100 seconds total

Question

Is this level of slowdown on Metal expected?
Are there known limitations or recommended optimizations for Metal?