Turnstone Judging with Strix Halo NPU?

A user benchmark posted to a forum reports that running the FLUX.2-klein-4B image model on the Strix Halo NPU via the Halogen engine generated images in 9.6 seconds at 7.6 GB of memory, versus 13 seconds on the GPU through stable-diffusion.cpp Vulkan, while concurrent LLM throughput fell only 10–15% (55/64 tok/s) instead of collapsing to 10 tok/s and 9 tok/s across 1 and 4 sessions. The same test found Z-Image-Turbo Q8 and Q4_K on the GPU took 31 seconds per image at roughly 10.7 GB and 8 GB respectively with perfect text rendering, while FLUX.2-klein-4B on the NPU produced nearly correct text. The poster concluded that image generation can run at essentially no cost to inference speed, though text quality on the NPU was described as not there.

Welcome everyone. Just got curious of the NPU discussion and can chime in. my strix halo can run image generation much faster and with very low impact to the performance using an image model on the npu, the quality is not there when it comes to text, but all other things considered i like the results along side the qwen 3.8 flash next on halogen: | | Model, engine | Time per image | Memory | Text in image | LLM while generating 1 / 4 sessions | |---|---|---|---|---|---| | 1 | Z-Image-Turbo Q8, stable-diffusion.cpp Vulkan GPU | 31 s | ~10.7 GB | perfect | collapses 64→10 / 75→9 tok/s | | 2 | Z-Image-Turbo Q4 K, sd.cpp Vulkan GPU | 31 s | ~8 GB | perfect | collapses same | | 3 | FLUX.2-klein-4B Q8, sd.cpp Vulkan GPU | 13 s | ~7.9 GB | garbled | GPU: same problem | | 4 | FLUX.2-klein-4B, Halogen on the NPU production | 9.6 s | 7.6 GB | nearly correct | 55 / 64 tok/s −10–15 % | So people can generate images to basically no cost of inference speed. I like it.