The Inner Workings of Tensor G4 in Pixel 9 Enlists DeepMind Team to Assist in Running Full-fledged AI Models

At the Pixel 9 launch event, Google introduced their new custom-designed Tensor G4 chip, a collaboration between DeepMind and the Google Silicon team. The focus was on optimizing artificial intelligence models for peak performance.

Tom’s Hardware had the opportunity to interview representatives from DeepMind and Google Silicon to delve deeper into this development. The goal of the Pixel 9 is to be the first smartphone capable of running the new Gemini Nano with Multi-Modality model, which can process inputs from various sources like images, sound, and text.

This challenge required the chip development team to respond to the demand by fine-tuning the chip and models simultaneously. The result was a threefold increase in processing capabilities, with an output rate of 45 tokens per second, the highest for a mobile chip. This also led to a 26% energy savings compared to previous models.

By combining enhanced AI processing capabilities with increased memory (12GB for standard models and 16GB for Pro models), the Pixel 9 can efficiently run high-powered AI tasks on-device. Features like Pixel Screenshots operate seamlessly without compromising user data privacy as processing is done locally.

The AI model running on Pixel 9 is a collaborative effort of Google’s internal teams – Android, Pixel, Google Silicon, and DeepMind. Apart from AI performance improvements, general user experience benefits include a 20% faster web browsing speed, 17% quicker app launches, and 20% longer battery life.

Google Silicon emphasized the challenges of benchmarking AI performance accurately, highlighting the disparity between academic benchmarks like MMLU and real-world applications. While there’s a push to optimize chips and models for high MMLU scores, it may not always reflect practical usage scenarios.

The speculated specifications of the Tensor G4 chip, as unofficially compiled by Android Authority, show a standard Arm core configuration: Cortex-X4 x1 + Cortex-A720 x3 + Cortex-A520 x4 (totaling 8 cores, one less than Tensor G3’s 9 cores). The GPU is expected to be the existing Mali-G715 from Tensor G3.

TLDR:
Google unveiled the customized Tensor G4 chip for Pixel 9, developed through collaboration between DeepMind and the Google Silicon team. The Pixel 9 aims to revolutionize AI processing with the ability to run the new Gemini Nano with Multi-Modality model efficiently, leading to improved performance and energy efficiency.