Google showed us a slew of new devices at Google I/O 2023, marking the official launch of the Google Pixel 7a, the Pixel Tablet, and the Pixel Fold. These devices join Google’s excellent existing lineup, with the Pixel 7 and Pixel 7 Pro currently spearheading it. Other than these devices being part of Google’s latest lineup, they have another common denominator: They are powered by the Google Tensor G2 chipset.

ANDROID POLICE VIDEO OF THE DAYSCROLL TO CONTINUE WITH CONTENT

But what makes Google’s second-generation chipset so special? And are there any downsides to Google’s custom processor compared to the competition? Read on to find out.


Google Tensor G2 specifications

For those that want a table, here you go:

Dev board code name

Cloudripper

Model number

GS201, Tensor G2

Cores

2x super-big ARM Cortex-X1, 2x big A78, 4x small Cortex-A55

GPU

Mali-G710 MP7

Manufacturing node

4nm Samsung PLP

Modem

Samsung Exynos 5300 5G

RAM

LPDDR5

Caches

4MB CPU L3, 8MP system level

The Tensor G2 is made by Samsung on its 4nm node using panel-level packaging. This is a complicated way of saying the chips are carved out of a square wafer rather than a round one, reducing waste. This likely doesn’t impact the chip’s performance much in actual devices, but it’s nifty and might reduce costs. Plus, it’s useful at a time when the industry is plagued by a chip shortage followed by a looming recession.

The Tensor G2 keeps the 2+2+4 core cluster configuration that the original Tensor GS101 used, with two “super-big” cores, two more typical big cores, and four small cores. One thing that changes across generations is the frequency and one small tweak to the big clusters. The A76 cluster is replaced by an A78 cluster that’s 100MHz faster at 2.35GHz. The other components remain the same. The X1 cluster has been bumped up by 50MHz, which gives it a frequency of 2.85GHz. This translates to a 10% to 15% better result in Geekbench, though, as with any such benchmarks, you’ll be hard-pressed to notice much of this difference in real life.

Google significantly upgraded the GPU. The Tensor G2 switched to the Mali-G710 GPU rather than the G78. That provides about 20% better performance and efficiency. The new GPU also helps the onboard machine-learning-focused Tensor Processing Unit (TPU), giving it an up to 35% boost in applicable processes. The TPU was also upgraded, with Google saying that it is up to 60% faster than its predecessor when it comes to photography and speech-related workloads.

The G2 is again paired with a Samsung-made modem, this time around, the Exynos S5300 5G. Mobile connectivity was poor on the Pixel 6 series and one of the biggest gripes many owners had with it. Based on our reviews and reports from Pixel 7 users, the situation is improved with the new modem. There are still some reports from people saying they have poor connectivity compared to other devices. Still, these seem to be few and far between, which could point to faulty devices or an area with poor connectivity.

This small upgrade might be disappointing on paper compared to the first-gen Tensor, but it could make a lot of sense in the performance-to-power usage ratio. Newer processors are found to improve performance at the cost of energy consumption, so sticking with the older generation cores might leave more room for better efficiency. It also helps that Google has experience with this setup for a whole generation, making it simpler to optimize the system further. This is somewhat reminiscent of the company sticking with the same camera for multiple generations of Pixel phones, improving how the software interacts with the hardware with each iteration.

According to one report, mass manufacturing for the chipset is estimated to begin as early as June 2022. This makes sense, as Google had to get enough Tensor G2 units ready for the Pixel 7 and 7 Pro launch on October 6, 2022, and needs a constant flow of new chips for its full lineup.

Google Tensor G2 special features

Smartphone chipsets aren’t just a list of cores. Other details inside them can affect performance. One of the biggest reasons that Google elected to create its own chipset with the original Tensor was for enhanced machine learning applications. With the rise of ambient computing, so-called heterogeneous compute (pushing specialized workloads to different or customized pieces of hardware rather than only general-purpose CPUs) has a bigger impact on perceived device performance than big single-threaded gains.

It’s not only about one or two big benchmarks. It’s also about how we use our phones. More and more, that’s for speech recognition, translation, fancy camera features, AR/VR, and other highly specialized workflows. And for that, you need more than a handful of recent ARM cores and a GPU.

The original Tensor included parts of Google’s HDRNet image processing pipeline in hardware, providing more specialization and direct performance for Google’s workloads than a general-purpose ISP can provide. Google also gave it a dedicated security core (paired with a separate Titan M2 chip running Trusty OS).

The Tensor G2 enables new such capabilities once again. At Google I/O 2022, Google talked about the Tensor G2 and said it would bring “even more AI-heavy breakthroughs and helpful, personalized experiences across speech, photography, video, and security.” And sure enough, the company delivered. The Pixel 7 can take and process night sight images up to two times faster than the Pixel 6. There is also a new Unblur feature on board that fixes blurred images. Further, speech recognition has been improved, with the Pixel 7 processing dictated text faster than the Pixel 6 without sending your audio snippets to servers.

Features like these are why a company would create a custom smartphone chipset. Otherwise, Google could have used something from Qualcomm, as it did in the past.

A render of the Google Tensor G2 chipset

While both the original GS101 Tensor and the Tensor G2 are mostly based on Samsung’s Exynos designs, we might see future models, like Tensor G3 (if it will be called that), diverge from this base over time and as Google’s requirements change.

It’s also worth pointing out that Google’s 2+2+4 core configuration is unique. So far, other chipset manufacturers haven’t followed in Google’s footsteps by including more than one “super-big” core. In an interview with Ars Technica’s Ron Amadeo, Google’s Phil Carmack (VP and GM of Google Silicon) said this configuration was chosen to increase efficiency at “medium” workloads by being able to throw more resources at a task to do it quickly, returning to a low-power state faster:

“When it’s a steady-state problem where, say, the CPU has a lighter load, but it’s still modestly significant, you’ll have the dual X1s running, and at that performance level, that will be the most efficient … You might use the two X1s dialed down in frequency so they’re ultra-efficient, but they’re still at a workload that’s pretty heavy. A workload that you normally would have done with dual A76s, maxed out, is now barely tapping the gas with dual X1s.”

Google Tensor G2-supported devices

The Tensor G2 debuted with the Pixel 7 series of phones, which includes the Google Pixel 7, Pixel 7 Pro, and Pixel 7a. The Pixel 7a and the Pixel 7, in particular, are similar phones, offering near-identical screens, camera setups, and RAM and storage. The Pixel 7 Pro is a little different in that it offers a much bigger screen with a higher refresh rate and an extra zoom camera that allows you to zoom up to 5x optically.

Google Pixel 7a, Pixel 7, and Pixel 7 Pro lying face-down next to each other on dark gray textured surface in sunlight.