Affiliate links on Android Authority may earn us a commission. Learn more.
Google Tensor G2 chip: Everything you need to know
With the Pixel 7 and Pixel 7 Pro, we also got the successor to the Google Tensor chipset. Dubbed the Tensor G2, this is Google’s second foray into the world of semi-custom silicon development, built in conjunction with Samsung Semiconductor. The Tensor G2 powers many Google devices that released in the first half of 2023, including the Pixel Tablet and Pixel Fold.
The original Google Tensor didn’t top any benchmarks, eschewing peak performance in favor of imaging, machine learning, and security silicon built to enhance specific user experiences. That design philosophy worked out reasonably well for the Pixel 6 series, albeit with some caveats regarding temperature and networking performance.
Google has taken the same semi-custom approach with the Tensor G2 too. At the core of the SoC, you’ll still find older and presumably cheaper off-the-shelf components that won’t be as snappy or efficient as the latest components on the market. Paired with Google’s next-gen custom AI and imaging silicon smarts, though, the Pixel 7 series banks on raw power being far less important. Instead, the focus remains on specially crafted hardware for Google’s bespoke speech, contextual support, imaging and video, and AI experiences.
Did Google get things right in its second attempt? Let’s take a closer look at what’s going on inside the Tensor G2 and what you can expect from it.
Editor’s note: Google’s Tensor G3 SoC has replaced the Tensor G2 as of the Pixel 8 series’ release. We have preserved the remainder of this article for historical context, with the exception of minor edits for relevancy.
Google Tensor vs Tensor G2 specs
Google Tensor G2 | Google Tensor | |
---|---|---|
CPU | Google Tensor G2 2x Arm Cortex-X1 (2.85GHz) 2x Arm Cortex-A78 (2.35GHz) 4x Arm Cortex-A55 (1.80GHz) | Google Tensor 2x Arm Cortex-X1 (2.80GHz) 2x Arm Cortex-A76 (2.25GHz) 4x Arm Cortex-A55 (1.80GHz) |
GPU | Google Tensor G2 Arm Mali-G710 MP7 | Google Tensor Arm Mali-G78 MP20 |
Caches | Google Tensor G2 4MB CPU L3 8MB system level | Google Tensor 4MB CPU L3 8MB system level |
RAM | Google Tensor G2 LPDDR5 | Google Tensor LPDDR5 |
Machine Learning | Google Tensor G2 Next-gen Tensor Processing Unit | Google Tensor Tensor Processing Unit |
Media Decode | Google Tensor G2 H.264, H.265, VP9, AV1 | Google Tensor H.264, H.265, VP9, AV1 |
Modem | Google Tensor G2 4G LTE 5G sub-6Ghz and mmWave | Google Tensor 4G LTE 5G sub-6Ghz and mmWave |
Process | Google Tensor G2 Samsung 5nm | Google Tensor Samsung 5nm |
First-gen Tensor vs. Tensor G2: What’s the difference?
As the table above highlights, there are only a handful of core changes between the original Tensor and the Tensor G2, and even they might not all be that meaningful.
For starters, the aging Cortex-A76 middle cores from 2018 have been replaced with 2020’s Cortex-A78. Following Arm’s claimed IPC improvements, these two cores offer more performance in exchange for marginally more area and power consumption. However, the other CPU cores remain virtually unchanged, with two powerful but several-year-old Cortex-X1s for the heavy lifting and four low-power Cortex-A55s for background tasks. We don’t get a wholesale performance uplift here; the overall CPU layout is primarily unchanged but provides a little more performance for gaming and other sustained workloads.
Google's Tensor G2 is a case of iterative rather than wholesale improvements.
The GPU layout is similarly revised yet not meaningfully different. Moving to Arm’s 2021 Mali-G710 microarchitecture offers a 20% performance and power improvement compared to the Mali-G78, and up to a 35% machine learning uplift. Impressive, and this may suggest in part why Google has switched from a beefy 20-core setup in 2021 to the smallest possible 7-core configuration. Although we should note that these new cores provide very different performance levels, so it’s not a direct comparison. We’ll discuss real-world performance results in the next section.
Google's upgraded TPU handles camera and speech tasks up to 60% faster.
Continuing the refinement trend, Google introduces its next-generation custom Tensor Processing Unit (TPU) inside the Tensor G2. Tightly coupled to the image pipeline, the TPU handles a wide range of machine learning tasks, from real-time translations to image and video processing. Google hasn’t specified exactly what’s new with the TPU but has told Android Authority that camera and speech tasks run up to 60% faster. Speaking of, the ISP now supports 10-bit HDR video recording, Google HDRnet tone mapping, and up to 108MP zero shutter lag images — though the Pixel 7 series has a 48MP camera.
Google Tensor G2 benchmarked: How does it perform?
With older CPU components and a comparatively compact GPU cluster, Google’s Tensor G2 was never going to hang at the top of the benchmark pack with the likes of the Qualcomm Snapdragon 8 Gen 2, MediaTek Dimensity 9200 series, and Apple A16 Bionic. And our Tensor G2 benchmarks show the deficit in no uncertain terms, even lagging behind previous-generation Snapdragon and MediaTek chips in some tests.
We see a 16% uplift in CPU performance going from the Pixel 6 to Pixel 7 in Geekbench 5. However, that’s still shy of the performance we saw in 2022 with the Snapdragon 8 Gen 1 and even further from the Gen 2 chip of 2023. As for GPU performance, it’s another lackluster showing on paper. We actually recorded a marginal win for the older Tensor chip in this area.
While these results may seem disappointing at first glance, it’s actually not all bad news. The improved efficiency we alluded to earlier allows the Pixel 7 devices to stretch their legs in real-life workloads for significantly longer than last year. Most users will care about this metric, as it has a direct impact on not just day-to-day use but also battery life in bursty workloads. Our Pixel 7 review noted similar or better battery life than last generation, even though Google reduced the battery capacity.
Tensor G2 vs. Snapdragon: How does it stack up against the competition?
As we’ve seen from benchmarks, chipsets from Apple and Qualcomm offer performance that far eclipses the original Tensor and the Tensor G2. Although Google’s chipset shares the same GPU cores as MediaTek’s very capable Dimensity 9000 Plus, the lower core count results in lower frame rates in high-performance games.
Moreover, Google told Android Authority that its Tensor G2 chipset is still based on Samsung Foundries’ 5nm process, which isn’t as efficient as the 4LPE node used by the Exynos 2200 and Snapdragon 8 Gen 1. TSMC’s N4 node, which Qualcomm first turned to in order to solve overheating for its 8 Plus Gen 1 model, is more efficient still. While Google may still have plenty of leverage in its custom machine learning silicon, its rivals aren’t standing still in this regard either — though the touted 60% performance boost should push Google out in front here.
Google Tensor G2 | Apple A16 Bionic | Snapdragon 8+ Gen 1 | Exynos 2200 | |
---|---|---|---|---|
CPU | Google Tensor G2 2x Arm Cortex-X1 (2.85GHz) 2x Arm Cortex-A78 (2.35GHz) 4x Arm Cortex-A55 (1.80GHz) | Apple A16 Bionic 2x Everest (3.46GHz) 4x Sawtooth (2.02GHz) | Snapdragon 8+ Gen 1 1x Cortex-X2 (3.2GHz) 3x Cortex-A710 (2.75GHz) 4x Cortex-A510 (2.0GHz) | Exynos 2200 1x Cortex-X2 (2.8GHz) 3x Cortex-A710 (2.52GHz) 4x Cortex-A510 (1.82GHz) |
GPU | Google Tensor G2 Arm Mali-G710 MP7 | Apple A16 Bionic Apple 5-core GPU | Snapdragon 8+ Gen 1 Adreno 730 | Exynos 2200 Xclipse 920 |
Caches | Google Tensor G2 4MB shared L3 8MB system level | Apple A16 Bionic 24MB system cache | Snapdragon 8+ Gen 1 6MB shared L3 4MB system level | Exynos 2200 Unknown |
RAM | Google Tensor G2 LPDDR5 | Apple A16 Bionic LPDDR5 | Snapdragon 8+ Gen 1 LPDDR5 | Exynos 2200 LPDDR5 |
Machine Learning | Google Tensor G2 Next-gen Tensor Processing Unit | Apple A16 Bionic 16-core Neural Engine | Snapdragon 8+ Gen 1 Hexagon | Exynos 2200 Dual-core NPU |
Media Decode | Google Tensor G2 H.264, H.265, VP9, AV1 | Apple A16 Bionic H.264, H.265, VP9 | Snapdragon 8+ Gen 1 H.264, H.265, VP9 | Exynos 2200 H.264, H.265, VP9, AV1 |
Modem | Google Tensor G2 4G LTE 5G sub-6Ghz and mmWave | Apple A16 Bionic 4G LTE 5G sub-6Ghz and mmWave | Snapdragon 8+ Gen 1 4G LTE 5G sub-6Ghz and mmWave | Exynos 2200 4G LTE 5G sub-6Ghz and mmWave |
Process | Google Tensor G2 Samsung 5nm (5LPE?) | Apple A16 Bionic TSMC N4 | Snapdragon 8+ Gen 1 TSMC N4 | Exynos 2200 Samsung 4LPE |
And just months after the Tensor G2’s release, the best Android phones of the year like the Galaxy S23 Ultra were powered by the Snapdragon 8 Gen 2. On the other hand, Google did not move to the latest available Arm architecture, which includes the powerhouse Cortex-X3 and Cortex-A715 CPUs. Luckily, the Pixel 7 series might not feel too far behind thanks to diminishing yearly improvements in the high-end chip space.
Put simply, we don’t have any complaints about the performance of the Tensor G2 in day-to-day tasks, and that’s not going to change unless some more demanding apps suddenly come along. Despite technology dating back to 2020, dual Cortex-X1 CPUs still offer more CPU grunt than you’ll need in most situations. And the increased efficiency and battery life of this generation means that we’re still getting a net upgrade over 2021’s chipset.
FAQs
The Tensor G2 delivers near flagship-grade performance for day-to-day smartphone tasks. However, it’s not the best chip for long gaming sessions.
No. The Tensor G2 outperforms some phones powered by previous-gen Snapdragon chips but it doesn’t match the 2023 Snapdragon 8 Gen 2’s performance.
Google’s Tensor G2 chip uses Arm CPU and GPU cores just like any other Android chip. However, the core configuration is unique, which affects performance.
Yes, Samsung Semiconductor fabricates the Tensor G2 chips for Google after the designing stage.
The Pixel 7 series, Pixel 7a, Pixel Tablet, and Pixel Fold all have a Tensor G2 chipset.