Google has just announced the Pixel 10 series, and all four phones are powered by the Tensor G5 processor. Tensor chips have traditionally enabled a variety of great Pixel-only features, but have also lagged behind rivals when it comes to peak horsepower and efficiency.
Nevertheless, the Tensor G5 is a landmark processor for the company as it’s the first Tensor chip made by TSMC instead of long-time partner Samsung. So what do we know about this silicon?
A cooler, more frugal Pixel chip (in theory) For starters, the Tensor G5 is made by TSMC on its 3nm process. Google didn’t identify the specific process used, but this switch from Samsung to TSMC is still good news for Pixel owners.
TSMC has traditionally been more adept at chip manufacturing than Samsung. This previously resulted in chips that were more efficient and powerful than the same chip produced by Samsung. Google hasn’t detailed specific efficiency improvements over the Samsung-made Tensor G4.
For what it’s worth, Google says the Pixel 10 phones enjoy “30+ hours” of battery life compared to the Pixel 9 range’s “24+ hours.” This is at least partially explained by the new phones’ slightly larger batteries, and we’re guessing other components are more efficient too. But the Tensor G5 will likely play a significant role here, too.
Many previous Pixel phones have also been plagued by heating issues, even when performing less demanding scenarios and tasks. The good news is that Google says it’s upgraded the “hardware and software thermal controls” to allow the chipset to run at higher clock speeds without throttling. Google specifically told Android Authority that the standard Pixel 10 has a “graphite thermal solution” while the Pro models have vapor chamber cooling. There’s no word if the vapor chambers are larger this time. We’ll just have to wait and see if these measures bear fruit or if heating problems persist.
Connectivity was another issue that plagued early Tensor-powered Pixels. Google seemingly got on top of the issues last year thanks to the new Exynos Modem 5400. So what’s new this time?
“Building on last year’s new modem, we are further optimizing this experience with software refinements,” the company told us, suggesting that the Exynos Modem 5400 has been carried over from the Pixel 9 family.
Tensor G5 CPU and GPU: Why so quiet, Google?
Rita El Khoury / Android Authority
The Tensor G5 offers a revised CPU layout compared to the Tensor G4. Google only noted during press briefings that the CPU consists of one big core, five medium cores, and two little cores. This would be a significant layout change over last year’s chip, which had one big core, three medium cores, and four little cores.
The company didn’t name the specific cores used or the clock speed. Nevertheless, Google later confirmed to us that the CPU has been upgraded to the “latest generation midcore IPs.” This suggests that the Tensor G5 is using Cortex-A725 cores at the very least. This would partially line up with our previous exclusive leak, which revealed one Cortex-X4 core, five Cortex-A725 cores, and two Cortex-A520 cores.
So what does this actually mean for CPU performance? Google claims the CPU is 34% faster on average than the Pixel 9’s chip. The company didn’t clarify whether this figure applied to single- or multi-core performance, although I expect much better multi-core performance thanks to the newer, more numerous middle cores. However, I’m not expecting the Tensor G5 to beat the Snapdragon 8 Elite chipset and Qualcomm’s custom Kryo cores.
The Tensor G5 enjoys a welcome multi-core CPU performance boost, but Google is suspiciously quiet on the GPU front.
What about the GPU, though? Google didn’t reveal any concrete information about the Tensor G5 GPU during its briefings, merely noting that it features “updated” GPU IP. The company also claimed that the Pixel 10 series ran many top games “very very well” during internal testing. Nevertheless, these vague statements aren’t a good sign if you were expecting a performance or efficiency upgrade.
Our previous Tensor G5 leak points to an Imagination DXT-48-1536 GPU, but we’ll need to confirm this with review units. That would be a major departure from Arm’s Mali GPU technology, which has been used since the first Tensor processor. Google also told us the Tensor G5 GPU doesn’t support ray tracing.
Qualcomm’s Adreno GPUs and Arm’s Mali graphics rule the roost as far as gaming and emulator support goes. So don’t be surprised if your Pixel 10 phone has compatibility issues with a few games and apps.
AI and imaging
Rita El Khoury / Android Authority
AI has become a key focal point for smartphone processors, so what does Google offer in this regard? The company claims that the Tensor G5’s TPU is up to 60% more powerful. Google says these gains are the result of architectural upgrades, such as more compute blocks, and a higher frequency. The nebulous nature of AI benchmarking means it’s still a little tough to tell how the Tensor G5’s AI horsepower compares to rival chips.
Nevertheless, Google says the Pixel 10 series supports a so-called Matryoshka transformer model, in reference to Russian nesting dolls. In any event, this architecture enables a smaller, speed-optimized AI model (two billion “effective” parameters) to be nested within a larger, quality-focused model (four billion “effective” parameters). The company says this approach means the phone doesn’t need two separate models, allowing apps to “dynamically” choose between the full model and sub-model.
The Tensor G5 has supercharged AI hardware, and supports some interesting AI model tricks.
Interestingly, Google explained to us that the submodel actually has five billion parameters while the full model actually has eight billion parameters. However, some of these parameters aren’t loaded into RAM. Instead, the Pixel maker uses per-layer embedding tech to pull these parameters in from flash storage in “tiny increments.”
So what do these improvements mean for AI performance? Well, Google claims that the latest Gemini Nano model running on the Tensor G5 is 2.6x faster and twice as efficient as the Pixel 9’s Tensor G4 for tasks like Pixel Screenshots and the Recorder app. The search giant also claims that the token window (how much input the model can process at once) has increased from 12,000 on the Tensor G4 to 32,000 on the new chip. Google says this larger token window is broadly equivalent to 100 screenshots or a month’s worth of emails.
What do you think of the Pixel 10's chipset so far? 57 votes It's looking good 25 % It seems okay 46 % It's looking bad 30 %
The improved AI silicon enables “over 20 unique experiences” at launch, with Google saying there’s more to come. Some of the more notable on-device features include 100x Pro Res Zoom, voice translations during phone calls (in the person’s voice), a Personal Journal app, Magic Cue, and more.
The Tensor G5’s image signal processor (ISP) has also been redesigned and works more closely with the digital signal processor. MediaTek and Qualcomm have taken similar approaches over the last few generations, drawing their ISPs and AI silicon closer together.
Google claims the chipset supports “advanced” segmentation to identify objects and details in a scene. It gave the example of segmentation being used in a portrait shot to identify skin, lips, eyes, and more for a more accurate image. Again, this isn’t anything new, as rivals have long offered advanced and even real-time scene segmentation in photos and videos. Nevertheless, I’m glad to see Pixels with these upgraded capabilities.
Google’s new ISP also enables a couple of video-related improvements, namely reduced blur in low-light videos, and 10-bit HDR by default for 1080p and 4K/30fps video capture. Unfortunately, you still have to rely on the cloud-based Video Boost feature for 8K video capture as Google chooses to skip local 8K recording on its phones.
Big questions remain about Tensor G5
Rita El Khoury / Android Authority
It’s hard to come to a firm conclusion about the Tensor G5 in the absence of some basic details. The company’s refusal to mention any GPU gains compared to the Pixel 9 series suggests that performance hasn’t changed, or that even a downgrade could be on the cards. We’ll also have to wait for review units to see whether the switch to TSMC has paid off in terms of battery life and if Google has definitively addressed heating problems.
One thing is for sure, though. Google’s Tensor processors have never really focused on peak performance or benchmarks. I’m not expecting this to change with the Tensor G5 and Pixel 10 series.
Follow