The math required to power AI on mobile devices is new, and Qualcomm is leading the way.
Stable Diffusion, an AI model for text-to image rendering that is typically run in the cloud using a web-browser and driven by servers in data centers with a lot of power and silicon horsepower, generated the feature image above. The image was created by Stable Diffusion on a smartphone running in airplane mode with no connection to the cloud data center. The AI model that rendered it was powered by Qualcomm Snapdragon 8 Gen 2, a mobile chip which operates on devices with less than 7 watts.
This image was rendered in 14.47 seconds by Stable Diffusion using only a few phrases.
This image is a scaled-up 540p input resolution that results in cleaner lines, sharper texture, and an overall better experience. Qualcomm offers a non-algorithmic variant of this today called Snapdragon GSR. However, in the future mobile enthusiasts gamers will be treated to better image quality, without sacrificing battery and frame rates.
This is only one example of gaming or media enhancement using pre-trained, quantized machine-learning models. But you can think of many other applications that would benefit, such as recommendation engines, location-aware navigation, computational photography techniques, and others.
Qualcomm appears to be leading the charge in implementing new maths for this AI heavy-lifting on smartphones and lower power edge devices.
Source:
https://www.forbes.com/sites/davealtavilla/2023/05/15/powering-ai-on-mobile-devices-requires-new-math-and-qualcomm-is-pioneering-it/?sh=f45f61e1a16d
