Repository navigation
Mac engine benchmark: share your Mac's report #43
Replies: 13 comments 3 replies
Kokon Mac engine bench
Where the time goes at 1024² (s): text encoder, DiT load, step 1, steps 2–8, VAE, total
Image 1 includes loading the models and building the Neural Engine's programs; image 2 is the engine already running, as in the app, 10 s after image 1 (below 32 GB the models are loaded again for every image). Accuracy at 512² against the engine's fp32 GPU path. |
Kokon Mac engine bench
Where the time goes at 1024² (s): text encoder, DiT load, step 1, steps 2–8, VAE, total
Image 1 includes loading the models and building the Neural Engine's programs; image 2 is the engine already running, as in the app, 10 s after image 1 (below 32 GB the models are loaded again for every image). Accuracy at 512² against the engine's fp32 GPU path. |
Kokon Mac engine bench
Where the time goes at 1024² (s): text encoder, DiT load, step 1, steps 2–8, VAE, total
Image 1 includes loading the models and building the Neural Engine's programs; image 2 is the engine already running, as in the app, 10 s after image 1 (below 32 GB the models are loaded again for every image). Accuracy at 512² against the engine's fp32 GPU path. |
|
TLDR: 4m 48s >> 2m 25s |
Kokon Mac engine bench
Where the time goes at 1024² (s): text encoder, DiT load, step 1, steps 2–8, VAE, total
Image 1 includes loading the models and building the Neural Engine's programs; image 2 is the engine already running, as in the app, 10 s after image 1 (below 32 GB the models are loaded again for every image). Accuracy at 512² against the engine's fp32 GPU path. |
Kokon Mac engine bench
Where the time goes at 1024² (s): text encoder, DiT load, step 1, steps 2–8, VAE, total
Image 1 includes loading the models and building the Neural Engine's programs; image 2 is the engine already running, as in the app, 10 s after image 1 (below 32 GB the models are loaded again for every image). Accuracy at 512² against the engine's fp32 GPU path. |
Kokon Mac engine bench
Where the time goes at 1024² (s): text encoder, DiT load, step 1, steps 2–8, VAE, total
Image 1 includes loading the models and building the Neural Engine's programs; image 2 is the engine already running, as in the app, 10 s after image 1 (below 32 GB the models are loaded again for every image). Accuracy at 512² against the engine's fp32 GPU path. |
|
Kokon 0.11.2 is out with Mac engine 0.1.1, and both of its main changes come from this thread. Thank you all.
Could you run the bench again on 0.11.2 and post the new report here?
On macOS 26 (Cesare, Alessandro), the report's "Paths selected" line should now say this Mac's signals arrive late. If it doesn't, please say so: that's worth knowing too. Thanks again for the time you put into this. |
Kokon Mac engine bench
Where the time goes at 1024² (s): text encoder, DiT load, step 1, steps 2–8, VAE, total
Image 1 includes loading the models and building the Neural Engine's programs; image 2 is the engine already running, as in the app, 10 s after image 1 (below 32 GB the models are loaded again for every image). Accuracy at 512² against the engine's fp32 GPU path. |
Kokon Mac engine bench v0.1.1
Where the time goes at 1024² (s): text encoder, DiT load, step 1, steps 2–8, VAE, total
Image 1 includes loading the models and building the Neural Engine's programs; image 2 is the engine already running, as in the app, 10 s after image 1 (below 32 GB the models are loaded again for every image). Accuracy at 512² against the engine's fp32 GPU path. |
Kokon Mac engine bench Kokon 0.11.2
Where the time goes at 1024² (s): text encoder, DiT load, step 1, steps 2–8, VAE, total
Image 1 includes loading the models and building the Neural Engine's programs; image 2 is the engine already running, as in the app, 10 s after image 1 (below 32 GB the models are loaded again for every image). Accuracy at 512² against the engine's fp32 GPU path. |
Kokon Mac engine bench
Where the time goes at 1024² (s): text encoder, DiT load, step 1, steps 2–8, VAE, total
Image 1 includes loading the models and building the Neural Engine's programs; image 2 is the engine already running, as in the app, 10 s after image 1 (below 32 GB the models are loaded again for every image). Accuracy at 512² against the engine's fp32 GPU path. |
Kokon Mac engine bench
Where the time goes at 1024² (s): text encoder, DiT load, step 1, steps 2–8, VAE, total
Image 1 includes loading the models and building the Neural Engine's programs; image 2 is the engine already running, as in the app, 10 s after image 1 (below 32 GB the models are loaded again for every image). Accuracy at 512² against the engine's fp32 GPU path. |
Uh oh!
There was an error while loading. Please reload this page.
Kokon's own engine for Apple Silicon (experimental) runs the GPU, the CPU's matrix unit and the Neural Engine together. Some of its paths are chosen per chip, and we can only measure the Macs we have (M2 Pro). Your report tells us whether those choices are right on your Mac.
How to run it
From the source instead:
git checkout macos-optimization, thenbash engine-mac/tools/bench.sh(writeskokon-bench-report.md).What the report has: chip, GPU cores, memory and macOS; time per step and per image at 1024×1024; peak memory; and how close the image is to the engine's most accurate (slow) path, with each chip-specific path on and off. A row with a path off that is faster than the default means that path should be off on your chip.
Nothing is sent anywhere by Kokon: the report is only what you paste. Add anything else that seems useful (other apps open, a laptop on battery, Low Power Mode).
All reactions