Google Tensor G3 and beyond

[limmk]

 

[limmk]

 

[limmk]

 

[limmk]

TL;DR

• Google’s in-house Tensor chips are facing heavy criticism from fans for their subpar performance and poor battery life.
• Readers feel that Google Pixels, which are priced as flagships, largely lack the performance to justify the price tag.
• While some fans appreciate the AI features and software integration of the chip, many believe that its flaws outweigh its benefits.

The recently launched Snapdragon 8 Elite Gen 5 looks promising on paper, and Qualcomm seems poised for another successful year, securing its place among top Android flagships once again. Despite its dominance, it’s looking unlikely that a Snapdragon SoC will make its way to a Google Pixel flagship anytime soon, and my colleague C. Scott Brown is actually glad that Pixels don’t use Snapdragon. Much like he predicted, this opinion was highly controversial, and it seems a lot of our readers are done forgiving Tensor for its mediocrity.

A recurring frustration that we could sense across over 130 comments in that article is that Tensor seems to have lost the plot for performance big time, especially when compared to Qualcomm’s Snapdragon and Apple’s A-series chips. The original article argued that not everyone needs blazing-fast performance, but readers are quick to point out that several of Google’s inefficient and outdated decisions on the Tensor are holding back the Pixel lineup, especially in areas such as gaming, heat management, and long-term performance.

Adding fuel to the Tensor fire is the fact that Pixels are priced like flagships, and yet, they don’t come with flagship performance. A few loyal fans do praise the tight software integration and AI features enabled by in-house chips, but the overall reader sentiment is heavily negative, with many likening the Tensor to a budget chip sold at luxury prices.

Take, for instance, our reader kylesbeautydiary, who mentions that it would be one thing if Tensor chips could approach today’s flagship performance, but they say that they don’t even come close to last year’s mid-range Snapdragon phones. The reader agrees that performance metrics aren’t everything, but at the same time, they do reflect an impact on everyday usage.

One reader recounted their personal experience upgrading from the Pixel 4 XL to the Pixel 8 Pro, only to find that some games that ran fine on the Pixel 4 were barely playable on the Pixel 8 Pro — one would presume that Tensor would be able to hold its ground against Snapdragon SoCs from four generations ago.

Reader Chaldon Pretorius presented several points against Tensor chips, namely that:

1. Pixels are worse value than competing Galaxy or iPhone models, and are blown away by the likes of OnePlus.
2. Pixels have significantly worse battery life than the competition.
3. Pixel cameras used to be great for photos, but they are just on par with the competition now as others have caught up. Pixels remain significantly behind in video.
4. Pixel’s value-adds are “AI slop” that ends up on other smartphones anyway.

The reader still agreed with the original premise of our article, that not everyone needs a “Porsche” and that Qualcomm needs competition. However, with that being said, users shouldn’t settle for “paying Porsche prices for a hopped-up Mustang.”

Reader Luke Vesty had a counter-argument to the last point, though. Just because Pixels ship with Tensor doesn’t mean that Google cannot charge a premium price, as a phone’s price is not defined by the chip alone. Pixels have numerous headlining features, both in hardware and software, that lend weight to its value proposition. Not everything is about benchmarks and battery drain tests; ultimately, it’s about the user experience.

One commenter even had a better Tensor-Snapdragon analogy than C. Scott’s Sprinter camper van-Porsche analogy:

Reader Jim Vlahos is a Pixel fan who prefers the software experience. However, even they agree that Snapdragon chips can do everything that Tensor can, but also do more, and do it faster and more importantly, more efficiently. The mediocre battery life is really getting old, and others agree that there is a big gap in battery life between Tensor and its competitors.

One thing is clear from the entire discussion: Tensor has a ton of room for improvement. Google has a lot of catching up to do, and much of the criticism coming its way is from fans who want to see Pixels succeed. Much like the company has to win back customers for its Home strategy, it also needs to win back smartphone fans with a home run.

 

[limmk]

 

[limmk]

Google initially debuted its Titan M security chip back in 2018 with the Pixel 3 before following it up with Titan M2 in its first-gen Tensor launch. Now, as we approach the tenth anniversary of Pixel and the five-year anniversary of that step into custom silicon, Google is rumored to be working on yet another step forward for its security coprocessor.

As reported by Mystic Leaks on Telegram, Google is apparently working on Titan M3 for this year’s Tensor G6, codenamed “Google Epic,” running firmware “longjing.” There’s not much else to go off of here other than these initial internal listings, though this leaked report suggests Google is aiming to compete more directly with Apple’s Secure Enclave. It’s certainly been a while since we’ve seen a fresh Titan M-series coprocessor, and whatever anniversary plans the company has in store for Pixel this year seems to suggest this is as good a time as any for a refresh.

But as for what to expect, all we can really do is speculate using past information. The blog post for Google’s initial Titan launch points to a handful of ways it’s designed to protect your phone, including bootloader validation, lock screen protection through a limited series of login attempts, and the ability to “generate and store” private keys through the then-new StrongBox KeyStore API launched with Android 9. When Google launched the RISC-V-based Titan M2 chip alongside Tensor, it promoted protection against “electromagnetic analysis, voltage glitching, and even laser fault injection.”

Google has certainly seen a handful of Pixel vulnerabilities over the last half-decade, though it’s been able to fix most (if not all) of them through monthly security patches. Whether Titan M3 brings improvements to that front — or, perhaps, goes for more out-there targets like the M2’s laser fault injection protection — remains to be seen. Either way, it’s shaping Tensor G6 up to be an interesting launch, and one that could help the brand during otherwise unsteady times.

 

[limmk]

 

[limmk]

The Pixel 11 series has left Google with something of a dilemma. Based on our time with the phones so far, they remain among the most enjoyable Android handsets to use, thanks largely to Google’s software experience and a growing collection of genuinely useful features. But our testing of the Tensor G6 shows that Google’s custom silicon is falling further behind rival flagship chips; early gaming tests suggest a familiar litany of issues; and a week with the Pixel 11 Pro suggests that many of Google’s newest AI tricks aren’t yet strong enough to make up the difference.

That’s not necessarily a problem for the Pixel 11, but it could become one for the Pixel 12.

Google doesn’t need to build the fastest smartphone chip in the world, and Tensor doesn’t need to win every benchmark to justify its existence. However, the Tensor G7 needs to close some of the enormous performance gaps, improve efficiency, and provide a hardware foundation for AI features that are actually worth using. Otherwise, Google risks reaching a point where the reasons to buy a Pixel increasingly rest on what the company did several years ago rather than what it’s doing today.

Tensor G7 needs a serious GPU upgrade​

The most obvious place for Tensor G7 to improve is performance. We saw some meaningful gains with the G6, particularly on the CPU and TPU side, but Google’s latest silicon remains well behind competing flagship processors. The GPU situation is even harder to defend, with rival chips offering more than twice the peak graphics performance in some tests.

Google can argue that benchmarks don’t represent everyday phone usage, and there’s some truth to that. A Pixel 11 Pro doesn’t feel twice as slow as a Samsung Galaxy S26 Ultra when opening apps or scrolling through social media. For most people, myself included, the Tensor G6 is more than powerful enough for the tasks they perform every day. But that doesn’t mean performance has stopped mattering.

Graphics support is a major weakness for the Tensor G6 that the next-gen needs to fix.​

CPU power remains important for demanding applications, gaming, video processing, multitasking, and increasingly sophisticated Android and desktop-class features. It’s also an important fallback for AI workloads that don’t run on Google’s TPU. I certainly don’t want to see Google stop improvements here, but the GPU is the component that has most troubled Tensor for the past few generations.

And it’s not just about benchmark numbers; Google’s choice of graphics determines how well the Pixel handles modern games and other GPU-bound computations, and how much performance headroom remains as those workloads become more demanding, such as running emulators for classic (and increasingly modern) consoles. Google’s choice of graphics silicon partner also has knock-on effects for driver stability and game support. All are areas where Google is now seriously lagging behind similarly priced rival phones.

That’s particularly important when Google promises seven years of software updates. A phone that’s already significantly behind its rivals at launch has less room to grow in the future. Tensor G7 doesn’t need to beat Snapdragon or Apple’s latest silicon across the board, but it does need to narrow the gap enough that Pixel buyers aren’t making an obvious long-term performance sacrifice just to get their hands on Google’s software today. Unfortunately, the Pixel 11 struggles to play the very best games today; it’ll be outright useless for the latest titles come the end of its seven-year lifespan. The Pixel 12 needs to drastically reevaluate that lack of value.
But this isn’t just about packing in more GPU power; it’s also about leveling the playing field for features. Moving to more budget-oriented PowerVR DXT and CXT parts has stagnated Tensor’s performance and feature set at a time when rivals are pushing ahead with ever more powerful rasterization, ray tracing, and AI upscaling features. Making the most of high-fidelity games and emulators on the go is just as much about power efficiency, quality drivers, and features ranging from ray tracing to variable rate shading as it is about sheer number crunching.

The Pixel 12 doesn’t need to become a gaming phone, but a flagship costing four figures should have enough graphics horsepower to play today’s demanding titles well while still leaving headroom for next-gen experiences. To do that, it might be time to revisit its GPU partnership with Arm or invest in something newer in the PowerVR portfolio. Whatever isn’t going to end up in another round of driver and support issues that take forever to fix.

TPUs need to make Google’s AI ambitions useful, or move over​

Of course, there’s a reason Google is willing to accept these compromises. Tensor was never designed solely to compete with Snapdragon on traditional performance metrics. Google wants its silicon to power the AI experiences that make a Pixel different from every other Android phone. The problem is that those experiences need to be good enough to justify the trade-off.

Our first week with the Pixel 11 Pro highlighted the issue. Many of Google’s new features are interesting, but several haven’t proven particularly useful in everyday life. HiLight is a clever piece of hardware, but it currently has very limited functionality. Proactive Assistance sounds promising, but it hasn’t consistently surfaced anything useful for us. Magic Capture and Camera Looks have potential, but neither has become something I would miss if I switched phones.

Meanwhile, some of the Pixel’s best features have been around for years. Now Playing, scam detection, At a Glance, and Pixel Screenshots remain excellent precisely because they quietly solve problems rather than constantly reminding you that your phone contains AI.

More TPU power is welcome, but it has to elevate the UX to be worth the tradeoffs.​

That’s an important lesson for the Pixel 12 in general, but also raises the question of whether Google is putting too much of Tensor’s limited silicon budget into the TPU for AI. Rival flagship chips already have dedicated AI accelerators of their own, but they’re not sacrificing CPU and GPU performance to lay them down on silicon. Qualcomm and MediaTek have long had AI acceleration across their SoCs, and newer chips now support Arm’s SME2 to run certain AI workloads efficiently on traditional CPUs. Likewise, modern GPUs (including mobile parts) increasingly support highly parallel quantized math, exactly the sort used for machine learning workloads, and mobile chipset architectures have shifted to shared-memory designs to quickly move data between CPU, NPU, DSP, and GPUs to make the most of their different capabilities.

Perhaps Tensor G7 should take a similar approach. Rather than simply making the TPU bigger and faster again, Google could invest in a broader range of AI acceleration options, giving the CPU and GPU greater ability to handle smaller, latency-sensitive workloads while reserving the TPU for the jobs it does best. That could also free up some of Tensor’s finite silicon budget for graphics, cache, memory subsystem, or efficiency improvements that Pixel users can actually feel outside of AI use cases.

Google’s AI strategy doesn’t need less ambition. But if the latest Pixel features aren’t proving transformative, I’m not convinced another huge TPU upgrade is necessarily the best way forward. The TPU needs to do more than look impressive on a spec sheet; it needs to enable experiences that rival phones simply can’t match. If it can’t, perhaps Tensor G7 would be better served by becoming a more balanced AI platform rather than an even more powerful TPU attached to a weaker flagship SoC.