Google’s Tensor chip goals to make the Pixel 6 smarter and last more

Google's Tensor chip aims to make the Pixel 6 smarter and last longer

Google has designed its personal processor for the primary time. 


Search big Google is taking a web page from Apple — by styling itself as a chip designer with extra management over its flagship smartphones. For homeowners of Google’s new Pixel 6, that would imply a cellphone with extra synthetic intelligence smarts that you will truly need to speak to and battery life lengthy sufficient to proceed the dialog. 

Google, which unveiled its $599 Pixel 6 and $899 Pixel 6 Professional telephones throughout a digital occasion Tuesday, launched adjustments like new colours, show enhancements, Android 12 and a horizontal digicam bump strip on the again as an alternative of a square-shaped nook array. However the greatest distinction is what’s inside: Tensor, the primary chip designed by Google.  

Phil Carmack, the vp and common supervisor in control of Google’s chip enterprise, and Monika Gupta, senior director of Google Silicon, chatted with CNET over video conferencing the week earlier than the Pixel occasion to get an in-depth take a look at what makes Tensor particular — and why Google even bothered to make the processor. 

Now playing:
Watch this:

Google unveils Tensor chip for Pixel 6 phones


“The key element in making this decision was around AI and how we could bring AI at a much different, more personal level to the end user,” Carmack said. “We simply weren’t able to get there with the existing solutions that were out there.”

The goal is to create a Pixel phone loaded with more AI smarts and the power to make those calculations without sacrificing battery life. In the new Pixel 6 and 6 Pro, that manifests through real-time language translations, highly accurate voice transcription and high-end camera features like the ability to unblur the face of a person in motion. On top of that, Google promises 24 to 48 hours of battery life. 

“It was kind of like we were being held back a little bit,” Gupta said. “We have access to state-of-the-art [machine learning] proper inside Google, but we could not deliver it to our Pixel customers. We’ve got this imaginative and prescient for Pixel, and we could not understand that imaginative and prescient.”

That adjustments with Tensor, which marks Google’s first foray into the costly and tough world of SoC design. It began engaged on the processor about 4 years in the past with a group of 76 individuals. (The group “is quite a bit larger now,” Carmack mentioned, although he declined to specify its dimension.) Most semiconductor corporations have hundreds of engineers creating their new chips. However Google has one thing else behind it — its group of researchers who focus on synthetic intelligence and machine studying. 

“On Tensor, we’re working the state-of-the-art [machine learning] fashions from Google Analysis, like the newest and biggest,” Gupta mentioned. “And we’re doing it way more effectively than ever doable earlier than as a result of Tensor structure.”

Whereas most handset makers use chips equipped by Qualcomm, Apple has stood out as an exception by creating processors for all the pieces from its iPhones to its Macs. Designing its personal processors lets Apple optimize for the options it cares about probably the most, like excessive efficiency and lengthy battery life. Now, Google is trying to comply with that technique because it faucets into its power in search and AI. The transfer offers a blow to Qualcomm.

“Qualcomm Applied sciences and Google have been companions for greater than 15 years, beginning with bringing the primary Android units to market,” Qualcomm mentioned in an announcement. “We are going to proceed to work intently with Google on present and future merchandise primarily based on Snapdragon platforms to ship the next-generation of consumer experiences for the 5G period.”

Google might use a recent course. Whereas its Android software program powers virtually 9 out of each 10 smartphones shipped globally, Pixels make up lower than 1% of telephones shipped around the globe, based on Technique Analytics. 

However Anisha Bhatia, a senior analyst at researcher GlobalData, known as Google’s efforts to make Tensor “important.”

“Making a chip is a posh and costly course of, and Google, an organization that is not sturdy in smartphones, is investing cash and sources into this course of,” Bhatia famous. “It might enable Google to raised compete within the smartphone market.”

Tensor’s specs

Tensor is what’s known as a system-on-a-chip, basically a monster processor that mixes the CPU — the brains powering the gadget — with different capabilities. 

Within the case of Google’s design, the eight-core CPU is built-in with a 20-core Arm Mali-G78 MP20 GPU, a Tensor Processing Unit machine studying engine, a sophisticated picture sign processor for images, a Tensor safety core, a “context hub” for ultra-low energy capabilities, and 8MB system cache. Tensor additionally features a 4MB CPU L3 cache and is constructed on the 5-nanometer manufacturing node. 

The secret’s all the elements of Tensor working collectively to hold out an motion on the Pixel 6. 

“It is form of uncommon that one in all these blocks is the star of any necessary expertise,” Carmack mentioned. “They must be rigorously choreographed.”

The Tensor safety core is a CPU-based subsystem that is separate from the primary apps processor. It permits delicate duties and controls to run in an remoted and safe surroundings. Alongside Tensor will sit a co-processor, Google’s next-generation, devoted safety chip known as Titan M2. And since Google created Tensor, it is extending safety assist to Pixel homeowners to 5 years, letting individuals maintain onto their units for much longer. 

Google Pixel 6 Pro

The Pixel 6 and Pixel 6 Professional (pictured right here) characteristic a digicam strip on the again as an alternative of a sq. digicam array within the nook.

James Martin/CNET

The CPU itself is eight cores primarily based on structure from chip designer Arm — two high-performance Cortex-X1 cores with clock speeds of two.8GHz, two midrange A76 cores at 2.25GHz and 4 small, high-efficiency A55 cores at 1.8GHz. Chip specialists like XDA Builders known as the configuration — which mixes Arm’s strongest new cores with older cores — uncommon when the specs leaked earlier than the occasion.

Carmack mentioned Google designed the CPU in a manner that might ship one of the best responsiveness and energy effectivity for intensive use instances like computational images.

“We did one thing completely different than the remainder of the Android ecosystem is doing as a result of we did not construct it as a way to win a benchmark of single threaded efficiency for the minimal quantity of {dollars} invested,” Carmack mentioned. “We constructed it to ship the experiences. … Having two actually high-performance cores helps us get higher total responsiveness and higher total sustained efficiency.”

Finally, the Pixel 6’s Tensor CPU efficiency is 80% sooner and its GPU is 370% sooner than the chip within the Pixel 5, Qualcomm’s midrange Snapdragon 765G. Notably, that chip was slower than the top-of-the-line processor present in different premium Android telephones

Nonetheless images

One of many areas that may see enhancements on the Pixel 6 is images. Already a powerful space for Pixels, Tensor allows even smarter digicam skills. 

A brand new characteristic known as Face Unblur places the face of a subject-in-motion in focus whereas sustaining the blur of the remainder of the physique. As an example, you’ll be able to seize a photograph of a kid leaping on a trampoline, with the face sharp and in focus and the limbs blurred in movement. 

The expertise works by tapping right into a Google Analysis face-detection ML mannequin known as FaceSSD. Earlier than even taking the photograph, the Pixel will detect there is a face within the scene. If it is blurry, the Pixel will routinely spin up the second digicam so it is able to go whenever you press the shutter button. 

The digicam takes two pictures concurrently, one from the primary digicam and one from the ultrawide digicam. The primary digicam makes use of regular publicity to present you a low-noise {photograph}, after which the ultrawide makes use of a sooner publicity to supply a brilliant sharp picture. 

“Machine studying will align these two pictures, it’s going to merge them, and it’ll take the sharper face from the ultrawide and the low-noise shot from the primary to get you kind of one of the best of each worlds,” Gupta mentioned. Lastly, the Pixel will take one closing look to see if there’s any blur remaining within the merged pictures. In that case, the cellphone will estimate the degrees and course of the blur and take away it. 

“All in all, it takes about 4 machine studying fashions to mix this knowledge from two completely different cameras after which ship you the {photograph} with a pleasant clear face,” Gupta mentioned. “Not solely is sort of the whole chip lit up with completely different subsystems doing various things, even on the cellphone stage, now we have a number of digicam sensors lit up. However this kind of will get to form of the guts of why we method Tensor otherwise and why our focus is de facto rooted in ranging from the tip customers.”

FaceSSD can even work with video, Gupta mentioned. The corporate has by no means been capable of apply face detection to video earlier than as a result of the older chips weren’t quick sufficient and consumed an excessive amount of energy, she mentioned. The brand new FaceSSD runs twice as quick, at 30 frames per second, which makes it appropriate for video. It is also decrease energy. 

“A variety of our computational images methods that we utilized to pictures traditionally, we at the moment are capable of apply the movies for the primary time,” Gupta mentioned. “All of those methods are simply going to make movies higher.”

One other images characteristic known as Movement Mode provides blur right into a nonetheless picture. It “brings kind of an expert feel and appear to city pictures, an evening out and even nature scenes,” Gupta mentioned. “Sometimes, you’d create these results with like perhaps a tripod or lengthy exposures or fancy gear and quite a lot of follow. However inside the Google Digicam app, we make this tremendous straightforward with the Movement Mode.”

The Pixel 6’s digicam takes a number of pictures and combines them. Utilizing the on-device machine studying and computational images, it’s going to establish the topic of the photograph, determine what’s transfer and what’s not, after which add a kind of aesthetic blur to the background. As an example, if taking a photograph of a bicycle owner, Movement Mode would make the bicycle owner in focus and blur the wheels and background barely to point the motion. 


Google Pixels have lengthy gained accolades for his or her photograph high quality, however video sometimes hasn’t measured as much as the standard present in rival telephones. The corporate goals to make its video simply pretty much as good as nonetheless pictures within the new Pixel 6 utilizing its computational images expertise, Gupta mentioned.  

Now playing:
Watch this:

First impressions of the new Pixel 6 and 6 Pro


Google developed an algorithm called HDRnet. It embedded parts of the algorithm directly into the Tensor image signal processor to “vastly” speed up the processing while “drastically” reducing the power consumption, she said. HDRnet allows the Pixel 6 to capture video at the quality of still images, and it can run on all video formats, even 4K video at 60 frames per second. 

“Once you put ML in the interactive loop, it’s a real-time guidance system now,” Carmack said. “An intelligent photographer is there in real time, making sure you get the right thing.”

Speech and transcription

Another area where Tensor benefits Pixel users is in how you talk to the device. The Pixel 6 comes with on-device speech recognition that “can transcribe speech with incredible accuracy,” Gupta said. Because of Tensor, that transcription consumes half the power as before. 

That improved speech recognition will help with features like Voice Access, which lets users talk to the phone without having to type in commands, and Google Call Screen, which uses Google Assistant to answer incoming calls, talk to the caller and provide a transcript of what the caller’s saying. 

Google Assistant also should have improvements, thanks to Tensor and the advanced ML enabled by the chip. The processor allows for more effective “hot word” detection — the “OK Google” that triggers the voice assistant to respond. In the Pixel 6, it can detect the hot word even when the background is noisy. 

Google Pixel 6 Pro

Google’s Tensor chip enables new features on the Pixel 6 Pro.

James Martin/CNET

And the new Tensor security core works with the rest of the processor to secure sensitive information. For instance, you may need to call on the Google Assistant to unlock your phone or grant access to your contacts. The Pixel 6 will do all of that processing on the device instead of sending it over the cloud. 

“On-device AI might help them placate some of the concerns people have around Google and better compete against iOS on the privacy and security side,” Creative Strategies analyst Carolina Milanesi said.

And because of Tensor, Google can apply more ML to problems to further refine and understand the nuances of the individual speaker, Gupta said. 

“This could be the difference between knowing is it Monica with the ‘C’ or Monica with the ‘K,’ or where do you put the comma, where do you put the punctuation, all the nuances you get depending on who’s speaking, based on their accent or based on the nouns they often use,” Gupta said. 

Live translation

When it comes to translation, the Pixel is running a new, “state-of-the-art” model called Neural Machine Translation, Gupta said. It consumes half the power than previously possible, Gupta said. 

Live Chat Translation quickly translates from one language to another. It works on any Android chat app, and translations happen directly in that app, rather than forcing a user to copy and paste text into Google Translate. The new Pixels also can translate media in real time, like providing live English interpretations on a French-speaking video. 

And thanks to Tensor and NMT, the new Pixel 6 can provide on-device translation for speech.

“Tensor was essentially designed to be the best on-device [machine learning] SoC out there,” Gupta mentioned. “If I am watching one thing, and it is being reside captioned for me or reside translated for me, that is non-public.”

These kinds of duties would sometimes drain a cellphone’s battery, if not for Tensor. 

“One in all our product targets was to form of remove this compromise with Tensor and actually push [performance and energy efficiency] concurrently,” Gupta mentioned. 

For Google, Tensor and the Pixel’s new AI options aren’t the tip. They’re only the start — they usually’re solely restricted by how highly effective and power environment friendly Google could make its chips.

“This tight collaboration between Analysis, Android platform, Pixel and the silicon group creates a pleasant virtuous cycle,” Carmack mentioned. “We’ve got an entire roadmap of merchandise.”

Leave a Reply

Your email address will not be published. Required fields are marked *

Related Posts