Google introduced its new chipset, the Tensor G5, alongside the Pixel 10 series. The Tensor G5 stands out as Google’s first processor produced on a 3-nanometer fabrication process. It’s also notable for being the first chipset produced by TSMC instead of Samsung. This switch offers significant gains in performance and efficiency.
Tensor G5 processor unveiled
The new chipset features a 1+5+2 core configuration, unlike last year’s Tensor G4, which features a 1+3+4 core layout. While clock speeds haven’t been officially announced, the Pixel 10 Pro Fold’s Geekbench 6 benchmark showed the most powerful core running at 3.78 GHz, the mid-range cores at 3.05 GHz, and the efficiency-focused cores at 2.25 GHz.
Google claims that the Tensor G5 is 34% faster than the Tensor G4 in average CPU performance. However, it hasn’t been stated whether this refers to single-core or multi-core performance. Despite this increase, the processor lags behind competitors like the Snapdragon 8 Elite, Dimensity 9400+, or A18 Pro.
Google didn’t provide any details about the GPU, only announcing that it uses an updated graphics processor. Previous reports suggested the chip would feature an Imagination Technologies DXT-48-1536 GPU, but this hasn’t been confirmed yet. The processor’s lack of ray tracing support hinders its performance, particularly in mobile games.
Google’s most prominent focus was AI performance. The new TPU is 60% more powerful than the Tensor G4. The chip features a new transformer model, Google called “Matryoshka,” with two billion effective parameters.
This makes the Gemini Nano 260% faster and twice as efficient as the previous generation. Google also stated that the token context window has been increased from 12,000 to 32,000, equivalent to 100 screenshots. Tensor G5 is responsible for more than 20 on-device AI features, including Magic Cue, Call Notes with action suggestions, the new Journal app, Fraud Detection, and Gboard Smart Edit.
For example, Voice Translation is powered by both a traditional voice machine learning model and a generative model working in parallel. By storing audio at once, Google can reproduce the speaker’s vocal characteristics with just a few seconds of live audio.
The camera has also been improved. Thanks to improvements in the ISP, blurriness in low-light videos has been significantly reduced. 10-bit HDR support is now available by default for 1080p and 4K 30fps video recording.
However, 8K native video recording is not available on the Pixel 10 series. Instead, users can use Google’s cloud-based Video Boost feature. Tensor G5 supports features such as Add Me, Auto Best Take, and 100x Pro Res Zoom in the camera.
100x Pro Res Zoom stands out as the first diffusion model to incorporate approximately 1 billion parameters and work within the app. While the TPU plays a critical role in this feature, the camera pipeline utilizes all components of the chip (CPU, DSP, ISP).
Google claims that the Pixel 10 series offers over 30 hours of battery life. This time was 24 hours for the Pixel 9 series. The new manufacturing process has significantly improved battery performance. The increase in battery capacity and the more efficient operation of other components also contribute to this increase.
{{user}} {{datetime}}
{{text}}