Leaks have advised us to lower our expectations for the Tensor G4, the chip that powers the Pixel 9. Google didn’t emphasize the chipset much during the launch event, which is quite telling. Post-launch discussions with three Google executives confirm that the leaks were accurate, but there’s no need for concern. In a conversation with The Financial Express, Pixel product manager Soniya Jobanputra explained that the Tensor G4 isn’t built for top-tier performance. This was echoed in an interview by Tom’s Guide with Silicon group product manager Jesse Seed and Google DeepMind product manager Zach Gleicher, who stated that the chipset is specifically optimized for the Pixel 9’s unique use cases.
When designing the chip, our focus isn’t on achieving high speeds or impressive feats. We’re not aiming to surpass any specific benchmarks. Instead, we’re designing it to fulfill our specific use cases. We recognized that opening apps was a pain point. Therefore, as we developed G4, we concentrated on improving that aspect to enhance the user experience.
Since the phones have not been released yet, we cannot compare their performance to previous models. Nonetheless, Jobanputra and Seed have noted that the chip enhances app launch speeds by 17 percent. Additionally, web performance has improved by 20 percent, and the user interface is now smoother and more responsive. Google has not disclosed the process node on which the Tensor G4 is built or its configuration details. According to rumors, the G4 was not the company’s initial choice for the series; however, Google opted for it upon realizing that a fully custom silicon solution would not be ready in time for the 2024 phone models.
We are also expected to see improvements in other areas since the chip has a better CPU and is more efficient.
When asked about gaming performance, Seed stated that both peak and sustained performance had improved.
Google is currently in its AI era, making it a focal point of discussion during the Pixel 9 launch. The Pixel team collaborated with Google’s AI-centric DeepMind research lab to optimize the device for running Gemini Nano with Multimodality, enhancing its ability to comprehend text, images, and audio. Although it may not be the quickest device, it boasts impressive capabilities like searching your gallery for a specific screenshot you can’t locate or seamlessly stitching together various photos.
We apply a similar strategy to our Tensor Processing Unit (TPU). We collaborate with the DeepMind team to predict the future trajectory of models and determine the types of models we aim to run on the device. Key considerations include the model’s size, the required processing speed, and potential bottlenecks such as memory bandwidth, which significantly impacts performance. These factors are especially critical when designing for use cases that utilize Gemini on-device models. We thoroughly consider all these elements during the chip design and development process.
In my opinion, our most significant innovation this year was becoming the first to use Gemini Nano with multi-modality on both silicon and a phone. This breakthrough enables some impressive use cases, such as Pixel Screenshots, which are extremely useful for recalling information.
The Tensor G4’s enhanced TPU enables it to achieve a peak output rate of 45 tokens per second, which is considered industry-leading. AI model performance is quantified in tokens per second, where tokens represent words or sub-words. The requests you submit are broken down into these tokens for processing. Additionally, Google’s Seed highlights that the Pixel 9 comes with more RAM compared to previous models, ensuring the AI model is always ready to assist whenever needed.
The Nubia Z60S Pro has arrived for review
The main point from these discussions is that Google is focused on specific use cases. Although the Pixel 9 may surpass other leading phones in terms of AI capabilities, it will be the least performant overall.
The initial indicators are not encouraging, as a leak suggests the CPU’s performance was throttled to 50 percent just two minutes into a stress test. However, this does not definitively indicate poor performance, since such tests are intended to push products to their limits.
Regardless, overheating has always been an issue with Pixel phones, and it would be unfortunate if this problem continued with the new series.
