The Pixel 10 series is powered by the Tensor G5, Google’s biggest upgrade to its custom silicon.
Google touts deeper customization, starting with how it’s manufactured by TSMC on the latest 3nm process node. This results in a CPU that is 34% faster on average, with “significant gains” in single and multi-threaded performance. Google says to expect improvements in responsiveness when browsing the web, launching apps, OS rendering, AI experiences, and other workloads.
There’s one big performance core, five mid-performance cores, and two efficiency cores. Google has also upgraded hardware and software thermal controls to let the chip operate at higher frequencies without throttling. Tensor G5 has a high-speed memory interface thanks to LPDDR5X (higher memory bandwidth) and UFS 4.0 (faster flash storage).
The TPU (Tensor Processing Unit) is up to 60% more powerful, while Tensor G5 runs the “newest” Gemini Nano model from DeepMind. This translates to Gemini Nano running “2.6x faster and 2x more efficiently for use cases like Pixel Screenshots and Recorder” compared to Tensor G4. There’s also a 32K token context window on G5 (which can be a month’s worth of emails or a hundred screenshots) compared to 12K last year.
The Tensor team touts collaboration with DeepMind on improving the quality of on-device models. G5 makes use of the Matformer Model Architecture to more efficiently run models, while Per Layer Embedding improves model response quality given the constraints of RAM on mobile devices.
Tensor G5 is responsible for over 20 on-device AI features at launch, like Magic Cue, Call Notes with actions, the new Journal app, Scam Detection, and Gboard Smart Edit. In the case of Voice Translate, the translation of what the speaker is saying involves a generative model running in parallel with a classical audio ML model. One-shot voice preservation means Google is able to reconstruct the speaker’s vocal characteristics from just a few seconds of live audio. On the privacy front, there’s no voice enrollment or audio being stored.
There’s also an upgraded custom Image Signal Processor that particularly helps video performance, especially in low-light conditions. The ISP performs more granular scene segmentation to understand regions and objects. That translates to the Pixel 10 recording 10-bit video by default at 1080p and 4K30. There’s also motion deblur and improved Real Tone.
On camera front, Tensor G5 powers Add Me, Auto Best Take, and 100x Pro Res Zoom. The latter is Google’s largest Pixel Camera model at nearly 1 billion parameters. It’s also the first diffusion model running in the app with the TPU helping play a big role, though the camera pipeline is using every part of the chip (CPU, DSP, ISP). For comparison, zoom on the Pixel 8 Pro’s G3 was tens of thousands of parameters, while Pixel 9 Pro’s’s G4 was in the low millions. (Another example of G5’s prowess is a 3-billion parameter multilingual model for real-time speech use cases.)
FTC: We use income earning auto affiliate links. More.
Comments