Five years ago, Google took a gamble. It built a custom system-on-chip, Tensor, designed not just to crunch numbers faster but to anchor the smartphone experience around machine learning. Back then, most people were still treating AI as a background buzzword. Today, with generative AI touching everything from photo editing to live translation, Google’s bet looks prescient.
“Tensor was the very first mobile SoC to anchor our mission around machine learning, and this has never been more relevant now that we’re fully in the Gemini era,” Jesse Seed, Group Product Manager, Silicon explained during the closed-door briefing.
Tensor as the “foundation” of Gemini runs through everything Google announced with the Pixel 10. The G5 chip isn’t just an annual upgrade in speed or efficiency. It’s Google’s largest leap yet. Built on TSMC’s 3 nm process, it brings a 34% faster CPU and a TPU with up to 60% more compute capacity while sipping 34% less energy. There’s new memory (LPDDR5X), new storage (UFS4), better thermal controls, and a redesigned ISP for imaging.
But raw specs only tell half the story. Tensor has always been about what AI can do once the silicon is tuned for it.
“Three years ago, the state of the art for camera zoom models was around 10,000 parameters,” Jesse explained. “This year, with Pixel 10 and Tensor G5, we’re running an in-camera diffusion model that’s nearly 1 billion parameters, on-device and in real time.”
Speech AI followed a similar trajectory. In 2021, Google’s ambition of running a billion-parameter on-device translation model was called a pipe dream. Today, Tensor G5 runs Gemini Nano with a staggering 3 billion parameters for real-time, multilingual speech processing.
Google says Pixel 10 launches with more than 20 Gemini Nano-powered features, like Voice Translate, which doesn’t just spit out subtitles but reconstructs your own voice in another language during a live call. Or Magic Cue, a proactive assistant that surfaces context before you even ask. Or Personal Journal, which generates writing prompts based on your day, and an upgraded scam detection system that quietly runs in the background.
Voice Translate on the new Pixel 10 series might be the best example. Speak into your Pixel 10, and Tensor G5 translates the call live, while preserving your voice without any robotic filters or an enrollment process. Just a few seconds of speech and suddenly you’re conversing across languages with your own vocal identity intact and it’s all happening locally on the phone.
“It’s not just about the complexity of the model,” Jesse said. “It’s about how often you can run it without draining your battery.”
Google claims over 30 hours of battery life despite the heavier AI load. That’s partly thanks to the efficiency gains at 3nm and partly to the TPU’s 34% boost in efficiency.
The secret sauce this year is how deeply Google’s silicon team worked with DeepMind, co-designing the latest version of their on-device Gemini Nano.
“For over a year, the Tensor team and DeepMind researchers have collaborated on this latest version,” Jesse said. “DeepMind would bring ideas to improve model quality. The Tensor team would feed back what made most sense for phones: memory, compute, power.”
Two standout innovations emerged from this collaboration. First, “MapFormer,” a Matryoshka-style transformer (“model within a model” to balance speed and quality) that nests a smaller sub-model inside a larger one, letting apps trade speed for quality on the fly. Second, per-layer embeddings (PLE), a trick that improves model responses without straining the memory.
The result: 2.6× faster and 2× more efficient AI tasks like screenshot summary or live transcription. And with a 32,000-token context window, triple than last year’s, Pixel 10 can juggle months’ worth of emails or 100 screenshots in one go.
Pixel phones live and die by their cameras and the Tensor G5 pushes it further. The new ISP works alongside the imaging DSP and TPU to unlock a 100× ProRes Zoom on the Pixel 10 Pro, up from 30× last year. To make that possible, Google built its first camera-specific diffusion model, trained with nearly a billion parameters, running directly inside the camera app.
“When you zoom beyond 30x, the model super resolves and recovers intricate details in seconds,” Jesse explained, showing details in an image taken from over 4.6 km away at the One World Trade Center in New York.
Alongside this AI magic, Google also stressed trust. The Pixel 10 camera now bakes in C2PA content credentials at the silicon level and every photo now carries cryptographic credentials that state whether it was edited or AI-touched.
Elsewhere, scene segmentation has grown smarter, enabling truer skin tones in harsh lighting. Low-light video blur has been reduced and HDR video at 1080p and 4K30 is now on by default. All of it feeds back into Pixel’s reputation for camera software that outpaces the hardware on paper.
Early results from Digit Test Labs give us the clearest picture yet of how the Tensor G5 stacks up against its predecessor.
On AnTuTu, the Pixel 10 Pro XL scored 1,321,483, a small 1.2% gain over the Pixel 9 Pro XL (1,306,281). But it lags far behind the Snapdragon 8 Elite (2,764,967) and Dimensity 9400 (2,359,676), roughly 52% and 44% lower, respectively.
In Geekbench single-core, performance sees a healthy jump: 2,153 vs 1,923 last year (+12%), but it still trails Qualcomm (3,065) and MediaTek (2,719) by 30% and 21%, respectively. Multi-core is flatter: the Pixel 10 Pro XL hit 4,557, just shy of the Pixel 9 Pro XL’s 4,601 (–0.9%). Against the Snapdragon 8 Elite (9,705) and Dimensity 9400 (8,090), Tensor G5 is well behind, by 53% and 44%, respectively.
Graphics is where Tensor finally flexes. On 3DMark Wildlife Extreme, the Pixel 10 Pro XL scored 3,240, up a strong 26% from the Pixel 9 Pro XL (2,562). Still, it’s almost half the performance of the Snapdragon 8 Elite (6,327) and Dimensity 9400 (6,161), 49% and 47% lower, respectively.
For everyday use and AI features, this might not matter. For gamers, it could. Of course, these are preliminary numbers and we will be running more detailed testing to validate and expand on these early findings.
Tensor G5 is the most convincing version of Google’s AI-first vision yet. The partnership with DeepMind, the billion-parameter camera models, the real-time translation, are not just party tricks, they’re glimpses of how AI-native phones might feel.
At the same time, the benchmarks remind us that Google is still not playing the same game as Qualcomm or MediaTek. The weakness is real, and pure performance isn’t Tensor’s story, but the experience is.
And maybe that’s the point. For Google, the win isn’t topping charts. It’s making AI feel invisible, helpful, and local. Tensor G5 underlines Google’s long bet that the future of smartphones is AI-first, everything else second.
