Tracking the Explosive World of Generative AI

Google Researchers Unleash AI Performance Breakthrough for Mobile Devices

A new breakthrough from Google AI researchers dramatically boosts the performance of generative AI models on mobile devices, moving us closer to a future where sophisticated AI can run on smartphones.

Google AI researchers have achieved blazing-fast results for running Stable Diffusion and other AI models on mobile phones. Photo illustration: Artisana

🧠 Stay Ahead of the Curve

  • Google AI researchers reveal a breakthrough method for accelerating generative AI models on mobile devices, with impressive results in their test of Stable Diffusion.

  • The breakthrough could enable complex generative AI models to run efficiently on smartphones, enhancing user experience and device capabilities.

  • A future where sophisticated AI applications become widely accessible and integrated into everyday handheld electronics is increasingly possible as performance improves.

By Michael Zhang

April 25, 2023

As generative AI models captivate users with their ability to create poems, images, and even videos, the computational resources needed to generate this content present a challenge for smaller devices. Consequently, many generative AI models are run on either cloud-based systems or high-powered computers.

Efficiently running generative AI models on mobile devices would mark a significant advancement in capabilities. In a recent paper, a team of Google AI researchers unveiled a method to dramatically improve the performance of the Stable Diffusion 1.4 model on mobile phones. One Samsung test device was able to generate images in under 12 seconds. Remarkably, this approach can be applied to other large diffusion models, potentially accelerating generative AI performance across a wide range of mobile devices.

The Breakthrough: A GPU-Optimized Shader

Central to the research breakthrough is a series of optimizations that take advantage of a mobile phone's Graphics Processing Unit (GPU). The researchers crafted a specialized shader, which is a program that the GPU runs to output rendered graphics, GPU-optimized shader, to execute multiple intermediate steps typically required by the Stable Diffusion model in a single step. Four GPU-specific optimizations into the shader, enabling improvements such as:

  • Executing multiple computations within a single draw call

  • Enhancing parallelism by performing calculations in blocks

  • Optimizing threading and memory cache management during shader execution

These combined optimizations yielded impressive results: overall image generation time decreased by 52% and 33% on a Samsung S23 Ultra and an iPhone 14 Pro, respectively, with memory usage also significantly reduced. As mobile phones typically have between 4 GB and 8 GB of RAM, efficient memory management is crucial for running generative AI models on smaller devices.

Researchers found substantial performance increases after layering in four optimizations.

The GPU optimizations demonstrate strong potential for the future of generative AI models on compact devices. While the Stable Diffusion model has around 1 billion parameters, other large diffusion models like OpenAI's DALL-E boast 3.5 billion parameters. Parameter count, which refers to the number of adjustable settings a model has to learn from its training dataset, generally correlates with the complexity and power of an AI model, as well as its computational demands.

Ultimately, the researchers emphasize that their breakthrough extends beyond merely enhancing the speed of the Stable Diffusion model. "These improvements expand the applicability of generative AI and elevate the overall user experience across a diverse range of devices," they conclude, hinting at a future where sophisticated AI models can operate swiftly on handheld electronics.

Read More: ChatGPT