Has Google’s Tensor venture failed?

Because of an unprecedented leak from Google’s gchips division, we already know nearly all the things concerning the processors coming in Google’s next-gen Pixel 10 and Pixel 11 flagships. Whereas Google’s Tensor venture has scored a couple of wins, most notably powering some distinctive AI and images options, 4 generations of chips have didn’t impress in key efficiency and energy effectivity metrics, leaving Google’s customers behind the curve. Sadly, in line with Google’s inner projections, the Tensor G5 and Tensor G6 will proceed to chart a nearly unchanged course.

Google’s subsequent two generations of smartphones will proceed to depart energy customers wanting in lots of regards. Would Google be higher off going again to Snapdragon? Has Google’s Tensor venture already failed? I’m afraid there are more and more robust arguments to be made right here.

Pixel followers have change into used to being removed from the chopping fringe of basic efficiency metrics. Google’s unique Tensor chip that debuted with the Pixel 6 collection arrived with dated CPU and GPU parts. The setup didn’t see main upgrades in these areas till the Tensor G3, and fairly a couple of vital parts, together with the GPU and TPU, have remained static with the newer Tensor G4. Whereas Tensor has clearly seen quite a few enhancements over the generations, they’ve been extra stop-start than a gradual stream of upgrades. Throw in doubtful battery life and modem connectivity points, significantly within the early generations, and Tensor’s historical past seems to be greater than a bit of shaky.

Proof, if extra had been wanted, might be seen within the canceled “redondo” model of the G4 — a extra formidable “absolutely customized” chip. As a substitute, we ended up with “zumapro,” primarily a slightly suped-up model of final 12 months’s “zuma” chip. Even so, these hoping that subsequent 12 months’s delayed breakaway from Samsung’s growth handholding would yield extra vital enhancements might be upset. As soon as once more, in line with Google’s leaked paperwork, the Tensor G5 ought to solely partially enhance the CPU setup, have a GPU improve that’s nonetheless mid-range (however could have ray tracing), and a modest 14% enchancment to the TPU for these unique Pixel AI options. At the least the chip might be extra environment friendly, because of TSMC’s 3nm-class N3E manufacturing course of.

Nonetheless, with minor efficiency enhancements over what’ll quickly be three generations, it’s onerous to see the place the additional headroom to energy thrilling new options will come from. Google can’t crunch extra on-device machine studying workloads with out a extra highly effective TPU, CPU, or GPU, suggesting that any groundbreaking Pixel 10 options usually tend to depend on cloud infrastructure than safer on-device processing.

For comparability, Qualcomm boasts a forty five% peak NPU efficiency uplift with its Snapdragon 8 Elite, ever-tighter integration between that AI processing core and its picture processor, and options like video object elimination working solely on-device. Whereas gauging precisely the place these rivals stack up when it comes to AI capabilities is troublesome as a result of vary of workloads, it’s clear that telephones quickly hitting the market powered by Qualcomm’s newest chip could have an enormous lead in CPU and GPU duties and will shut in on and even overtake Google’s machine studying capabilities. It absolutely received’t be lengthy earlier than options comply with go well with. If that occurs, it’s important to query what advantages Tensor actually brings to the desk.

Sticking with Qualcomm for a second, cutting-edge efficiency doesn’t come low cost. Snapdragon costs are reported to have elevated sharply in recent times, peaking at round $150 for the Snapdragon 8 Elite. That’s an enormous chunk of a typical flagship’s $999 MSRP and is little doubt consuming into these already tight revenue margins. Subsequent-gen telephones are more likely to value extra consequently. Even Apple, which may make processors for iPhones principally at value and has the good thing about big gross sales quantity, is estimated to spend $100-$150 on its A18 Professional and outsourced energy and networking parts.

By comparability, we all know that Google is concentrating on round $65 for the price of the Pixel 11’s Tensor G6, far lower than what rivals might be shelling out on Qualcomm silicon in 2026 and seemingly simply two-thirds of Apple’s present spend (if these prices embrace all the things). There’s no manner Google can compete when it comes to efficiency and options in the long run if it might’t spend the same quantity on silicon growth.

Equally, manufacturing on bleeding edge nodes, mandatory to appreciate all-important vitality effectivity positive aspects, is turning into notoriously costly, extra in order TSMC races to next-gen 3nm nodes and under. Google’s Tensor G5 is predicted to be bigger than Apple’s present A18 Professional, so it would value extra to provide, at the least when it comes to silicon space. But, it would find yourself a great distance behind in lots of metrics.

To steadiness the books, Google is planning to take an axe to the Tensor G6’s silicon space, aiming to shrink it by some 8% over the G5. This might be completed by apparently yanking ray tracing from the GPU only a era after it arrived, the DSP will drop a core, and the system-level cache (vital for sharing knowledge between the CPU and peripherals) could be ditched. The G6 ought to debut new, quicker CPU cores, however the structure will shrink to only seven cores, lowering the influence of the improve. In different phrases, Google is making a acutely aware option to downgrade or not enhance elements of its chip to avoid wasting prices and prioritize area for AI and imaging — at the least in line with the leaked plans. All chips should make compromises, after all, however Google’s pretty low finances means it’s making harder decisions than flagship rivals.

Google estimates it’ll end in a modest CPU uplift and roughly the identical GPU efficiency as its predecessor, making the G6 one other half improve. Nevertheless, this would possibly change, because the paperwork we’ve seen had been most likely revealed earlier than the ultimate design was settled. Fortunately, Google’s docs recommend some new TPU and ISP improvements are on the best way; Google seems to be planning 100x zoom and “Extremely Low Mild video” options, in addition to basic picture high quality and bokeh enhancements and help for higher ML mannequin compression. There’s additionally point out of a brand new NanoTPU for well being and voice duties and an unspecified total TPU increase that must be bigger than the G5’s slightly meager 14% over the G3 and G4.

With prices on the forefront, the Tensor G6 seems to double down on Google’s dedication to AI and digital camera options over outright efficiency. We’ll simply should see if that’s sufficient to carry its rivals at bay, who may also be making their very own enhancements to images, AI, and uncooked efficiency over the following two years.

Pulling again to take a look at Google’s Tensor efforts in full, it appears that evidently its plans have fallen foul of the most typical pitfalls of customized silicon growth — it may be extraordinarily costly, and staying aggressive is fraught with potential setbacks. Lacking “redondo” appears to have skewed the entire venture’s trajectory considerably, however Google actually wouldn’t be the primary to fall delayed.

Samsung struggled for years with its customized Arm-based Mongoose cores in Exynos, usually failing to hit efficiency and effectivity targets. It will definitely ditched the venture and settled on utilizing Arm Cortex CPUs. Qualcomm additionally didn’t anticipate Apple’s fast change to 64-bit in 2013, inflicting it to drop customized CPU growth for Arm’s cores till the 8 Elite’s Oryon CPU revived the concept and is now inflicting its rivals complications. Customized growth of any sort is fraught with the potential for missed deadlines and being blindsided by competitor developments. Tensor has seemingly suffered greater than most in recent times, with stop-start enhancements and part freezes slightly than the regular year-on-year development we see from its rivals.

Likewise, if Google is immediately reeling from the prices of silicon growth, that means at the least some factor of poor long-term planning. We’ve identified that the advantages of smaller manufacturing nodes have been coming at more and more steep costs for years. Tensor has a big, costly silicon space but hasn’t used this area to leapfrog its opponents in any significant sense. Likewise, licensing the newest and best parts from Arm, Creativeness Applied sciences, and so on., will not be low cost (therefore the mid-range GPUs), neither is the price of self-developing TPUs, safety chips, and the like.

A top-class Tensor was at all times going to be an costly endeavor, however the funding wanted to be justified by clear wins. Benchmarks are only a small piece of the puzzle, after all, and after we have a look at real-world metrics like gaming efficiency and battery life, Tensor stays within the combine and has made clear enhancements over time. Nevertheless, when you think about that Pixel telephones are as costly as rival fashions, I’m unsure Tensor scores sufficient clear wins. It actually doesn’t completely personal the images or AI areas.

So, has Tensor failed? Properly, that could be a bit harsh, given the Pixel 9 is promoting nicely. Joe Public isn’t too bothered about specs, particularly when apps and video games run tremendous (even when not the quickest round), and battery life is now respectable. Google efficiently leverages its customized TPU design to construct eye-catching AI and digital camera options for its Pixel collection, which, if we imagine Google, is the important thing goal. It’ll additionally construct on these options within the subsequent few years, significantly when the Tensor G6 and the Pixel 11 roll round.

Nevertheless, I don’t assume Tensor has been the success story that Google envisioned and might be more and more a growth headache slightly than an asset. Between finding out doubtful battery life, failing additional behind the competitors in conventional efficiency, and cost-cutting measures, Tensor’s previous and future are combined at greatest.

Maybe the mounting case in opposition to Tensor hinges on the truth that Google is charging greater than ever for the Pixel 9 collection, but AI and point-and-shoot photos are simply a part of the image on the $1,000 mark. Even when it’s an unfair notion, energy customers and avid gamers are lukewarm about what Google has to supply, and our present information suggests the Pixel 10 and 11 received’t deal with that. To not point out the truth that OK efficiency immediately received’t really feel like that in 5, six, or seven years. Tensor’s meandering growth cycle exposes an actual threat of rivals leaving it utterly behind in upcoming generations.

That mentioned, Tensor’s distinctive options undoubtedly assist the entry-level Pixel flagships and inexpensive A-series stand out, as efficiency issues are much less vital. As a substitute of working again to Snapdragon, maybe it’s time for Google to contemplate a two-tier chip technique.