Has Google’s Tensor project failed?

Thanks to an unprecedented leak from Google’s gchips division, we already know virtually everything about the processors coming in Google’s next-gen Pixel 10 and Pixel 11 flagships. While Google’s Tensor project has scored a few wins, most notably powering some unique AI and photography features, four generations of chips have failed to impress in key performance and power efficiency metrics, leaving Google’s consumers behind the curve. Unfortunately, according to Google’s internal projections, the Tensor G5 and Tensor G6 will continue to chart a virtually unchanged course.

Google’s next two generations of smartphones will continue to leave power users wanting in many regards. Would Google be better off going back to Snapdragon? Has Google’s Tensor project already failed? I’m afraid there are increasingly strong arguments to be made here.

Pixel fans have become used to being far from the cutting edge of classic performance metrics. Google’s original Tensor chip that debuted with the Pixel 6 series arrived with dated CPU and GPU components. The setup didn’t see major upgrades in these areas until the Tensor G3, and quite a few important components, including the GPU and TPU, have remained static with the newer Tensor G4. While Tensor has clearly seen numerous improvements over the generations, they’ve been more stop-start than a steady stream of upgrades. Throw in dubious battery life and modem connectivity issues, particularly in the early generations, and Tensor’s history looks more than a little shaky.

Proof, if more were needed, can be seen in the canceled “redondo” version of the G4 — a more ambitious “fully custom” chip. Instead, we ended up with “zumapro,” essentially a marginally suped-up version of last year’s “zuma” chip. Even so, those hoping that next year’s delayed breakaway from Samsung’s development handholding would yield more significant improvements will be disappointed. Once again, according to Google’s leaked documents, the Tensor G5 should only partially improve the CPU setup, have a GPU upgrade that’s still mid-range (but will have ray tracing), and a modest 14% improvement to the TPU for those exclusive Pixel AI features. At least the chip will be more efficient, thanks to TSMC’s 3nm-class N3E manufacturing process.

Still, with minor performance improvements over what’ll soon be three generations, it’s hard to see where the extra headroom to power exciting new features will come from. Google can’t crunch more on-device machine learning workloads without a more powerful TPU, CPU, or GPU, suggesting that any groundbreaking Pixel 10 features are more likely to rely on cloud infrastructure than more secure on-device processing.

For comparison, Qualcomm boasts a 45% peak NPU performance uplift with its Snapdragon 8 Elite, ever-tighter integration between that AI processing core and its image processor, and features like video object removal running entirely on-device. While gauging exactly where these rivals stack up in terms of AI capabilities is difficult due to the range of workloads, it’s clear that phones soon hitting the market powered by Qualcomm’s latest chip will have a huge lead in CPU and GPU tasks and could close in on or even overtake Google’s machine learning capabilities. It surely won’t be long before features follow suit. If that happens, you have to question what benefits Tensor really brings to the table.

Sticking with Qualcomm for a moment, cutting-edge performance doesn’t come cheap. Snapdragon prices are reported to have increased sharply in recent years, peaking at around $150 for the Snapdragon 8 Elite. That’s a huge chunk of a typical flagship’s $999 MSRP and is no doubt eating into those already tight profit margins. Next-gen phones are likely to cost more as a result. Even Apple, which can make processors for iPhones mostly at cost and has the benefit of huge sales volume, is estimated to spend $100-$150 on its A18 Pro and outsourced power and networking components.

By comparison, we know that Google is targeting around $65 for the cost of the Pixel 11’s Tensor G6, far less than what rivals will be shelling out on Qualcomm silicon in 2026 and seemingly just two-thirds of Apple’s current spend (if these costs include everything). There’s no way Google can compete in terms of performance and features in the long term if it can’t spend a similar amount on silicon development.

Equally, manufacturing on bleeding edge nodes, necessary to realize all-important energy efficiency gains, is becoming notoriously expensive, more so as TSMC races to next-gen 3nm nodes and below. Google’s Tensor G5 is expected to be larger than Apple’s current A18 Pro, so it will cost more to produce, at least in terms of silicon area. Yet, it will end up a long way behind in many metrics.

To balance the books, Google is planning to take an axe to the Tensor G6’s silicon area, aiming to shrink it by some 8% over the G5. This will be accomplished by apparently yanking ray tracing from the GPU just a generation after it arrived, the DSP will drop a core, and the system-level cache (important for sharing data between the CPU and peripherals) might be ditched. The G6 should debut new, faster CPU cores, but the layout will shrink to just seven cores, reducing the impact of the upgrade. In other words, Google is making a conscious choice to downgrade or not improve parts of its chip to save costs and prioritize space for AI and imaging — at least according to the leaked plans. All chips have to make compromises, of course, but Google’s fairly low budget means it’s making tougher choices than flagship rivals.

Google estimates it’ll result in a modest CPU uplift and roughly the same GPU performance as its predecessor, making the G6 another half upgrade. However, this might change, as the documents we’ve seen were probably published before the final design was settled. Thankfully, Google’s docs suggest some new TPU and ISP innovations are on the way; Google appears to be planning 100x zoom and “Ultra Low Light video” features, as well as general image quality and bokeh improvements and support for better ML model compression. There’s also mention of a new NanoTPU for health and voice tasks and an unspecified overall TPU boost that should be larger than the G5’s rather meager 14% over the G3 and G4.

With costs at the forefront, the Tensor G6 appears to double down on Google’s commitment to AI and camera features over outright performance. We’ll just have to see if that’s enough to hold its rivals at bay, who will also be making their own improvements to photography, AI, and raw performance over the next two years.

Pulling back to look at Google’s Tensor efforts in full, it seems that its plans have fallen foul of the most common pitfalls of custom silicon development — it can be extremely expensive, and staying competitive is fraught with potential setbacks. Missing “redondo” seems to have skewed the whole project’s trajectory somewhat, but Google certainly wouldn’t be the first to fall behind schedule.

Samsung struggled for years with its custom Arm-based Mongoose cores in Exynos, often failing to hit performance and efficiency targets. It eventually ditched the project and settled on using Arm Cortex CPUs. Qualcomm also failed to anticipate Apple’s quick switch to 64-bit in 2013, causing it to drop custom CPU development for Arm’s cores until the 8 Elite’s Oryon CPU revived the idea and is now causing its rivals headaches. Custom development of any kind is fraught with the potential for missed deadlines and being blindsided by competitor developments. Tensor has seemingly suffered more than most in recent years, with stop-start improvements and component freezes rather than the steady year-on-year progression we see from its rivals.

Likewise, if Google is suddenly reeling from the costs of silicon development, that suggests at least some element of poor long-term planning. We’ve known that the benefits of smaller manufacturing nodes have been coming at increasingly steep prices for years. Tensor has a large, expensive silicon area yet hasn’t used this space to leapfrog its competitors in any meaningful sense. Likewise, licensing the latest and greatest components from Arm, Imagination Technologies, etc., is not cheap (hence the mid-range GPUs), nor is the cost of self-developing TPUs, security chips, and the like.

A top-class Tensor was always going to be an expensive endeavor, but the investment needed to be justified by clear wins. Benchmarks are just a small piece of the puzzle, of course, and when we look at real-world metrics like gaming performance and battery life, Tensor remains in the mix and has made clear improvements over the years. However, when you consider that Pixel phones are as expensive as rival models, I’m not sure Tensor scores enough clear wins. It certainly doesn’t exclusively own the photography or AI spaces.

So, has Tensor failed? Well, that might be a bit harsh, given the Pixel 9 is selling well. Joe Public isn’t too bothered about specs, especially when apps and games run fine (even if not the fastest around), and battery life is now decent. Google successfully leverages its custom TPU design to build eye-catching AI and camera features for its Pixel series, which, if we believe Google, is the key objective. It’ll also build on these features in the next few years, particularly when the Tensor G6 and the Pixel 11 roll around.

However, I don’t think Tensor has been the success story that Google envisioned and is probably increasingly a development headache rather than an asset. Between sorting out dubious battery life, failing further behind the competition in traditional performance, and cost-cutting measures, Tensor’s past and future are mixed at best.

Perhaps the mounting case against Tensor hinges on the fact that Google is charging more than ever for the Pixel 9 series, yet AI and point-and-shoot pictures are just part of the picture at the $1,000 mark. Even if it’s an unfair perception, power users and gamers are lukewarm about what Google has to offer, and our current knowledge suggests the Pixel 10 and 11 won’t address that. Not to mention the fact that OK performance today won’t feel like that in five, six, or seven years. Tensor’s meandering development cycle exposes a real risk of rivals leaving it completely behind in upcoming generations.

That said, Tensor’s unique features undoubtedly help the entry-level Pixel flagships and affordable A-series stand out, as performance considerations are less important. Instead of running back to Snapdragon, perhaps it’s time for Google to consider a two-tier chip strategy.