⚡ The hard part of on-device AI isn't the model. It's the 16.6ms. At 60fps that's your entire bud...⚡ The hard part of on-device AI isn't the model. It's the 16.6ms. At 60fps that's your entire bud...
The network for creativity
Join 1.25M professional creatives like you
Connect with clients, get discovered, and run your business 100% commission-free
Creatives on Contra have earned over $150M and we are just getting started
Back to feedPost
Iftekher Toufique's avatar
Iftekher Toufique Imam
• 17h
⚡ The hard part of on-device AI isn't the model. It's the 16.6ms.
At 60fps that's your entire budget — capture, preprocess, inference, postprocess, render. Blow it and users don't say "the AI is slow." They say the app is broken.
What actually bought us headroom shipping photo editing apps:
• Keep tensors on the ANE — every CPU↔GPU copy costs more than the layer you were busy optimizing
• fp16 first, measure, then reach for int8
• Run inference at preview resolution; go full-res only on export
• Profile on the oldest device you support, not the one on your desk
Nobody claps for the architecture. They clap for the thing that felt instant.
ondeviceaiiOSmobiledeveloper
Post image
Back to feed
The network for creativity
Join 1.25M professional creatives like you
Connect with clients, get discovered, and run your business 100% commission-free
Creatives on Contra have earned over $150M and we are just getting started