What do you do when advancements in software outpace those in hardware? You just start building the hardware yourself.
Google said on Monday that its next flagship phones—the Pixel 6 and Pixel 6 Pro, both due this fall—will be powered by a custom-built processor called Tensor. Codesigned over the past four years with machine learning and artificial intelligence experts, it’s built for all the Android-based wizardry in the upcoming Pixel phones.
The move comes at a time when tech industry giants increasingly favor custom designs for the processors that power their phones, laptops, and other gadgets.
According to Google hardware chief Rick Osterloh, the company has hit the limit of what it can build with other companies' chips. “The problem always boils down to hardware capability,” Osterloh says. “Are you able to actually do the processing necessary to really run sophisticated advanced AI models? Unquestionably, you run into different constraints with off-the-shelf technology. So several years ago, we decided if we’re going to really innovate for the future, we’re going to need to build our own system."
When Osterloh says “several years ago,” he's referring to the first computer chip Google revealed back in 2016. Google calls the chip Tensor Processing Unit, or TPU, a nod to the company's open source machine learning platform TensorFlow. These TPUs are application-specific chips, designed to support machine learning. But they were also built for the AI servers that live inside giant data centers. This new Tensor chip will be right inside the Pixel device in your hand or pocket—at least, it will be whenever the Pixel 6 phones arrive in October.
“We knew we wanted to borrow on the equity of that name, and eventually run TensorFlow really efficiently on mobile devices,” Osterloh said. By building its own custom mobile chip, Google isn't just juicing its phones with Google-y artificial intelligence; it's joining the ranks of Apple and Samsung, which both design custom silicon for their top-of-the-line consumer electronics. It's going deeper into the complex world of the chip supply chain (while still maintaining a relationship with chipmaker Qualcomm for its lower-end Pixel phones, at least for now). And it's doing it right in the midst of a global chip shortage.
Chip 'n' Dip
Google has been churning out some of the best smartphone cameras for a few years now, and it did this with the help of powerful machine learning and AI models—what's known as computational photography, instead of jamming a bigger camera sensor into the phone.
It's software perks like these that are supposed to make Pixel phones stand out from the crowd of other Android phones, even if Google doesn't sell all that many Pixels. You might not get the fastest phone around, but a Pixel can, for example, screen an incoming call to see if it’s a robocall, passively identify what music is playing at the coffee shop, and live-transcribe videos on Instagram, even if you’re offline.
Until now, Pixel phones have been powered by Qualcomm processors, which power most (non-Apple) phones sold in the US. This new Tensor chip is still based on the same ARM architecture as Qualcomm and other mobile processors, so customers don't need to worry about losing compatibility with modern-day apps or anything of the sort. And Osterloh says Google's lower-end Pixel phones, which have an “A” attached to the name (such as the upcoming Pixel 5A) will still run on Qualcomm Snapdragon chips.
But by designing its own chip, Google is able to better optimize its hardware and software—precisely the reason why Apple developed its own processors instead of relying on Intel chips for its Macs. It’s powerful enough to handle intensive AI and ML models on device (without needing to rely on a remote cloud server). Osterloh didn’t disclose exactly who is manufacturing Tensor, but says, “We designed it, own it, and drive it. This is our chip.”
What exactly does this mean for the next wave of Pixels? Osterloh says pretty much every current Pixel feature powered by AI and machine learning—like Night Sight or Portrait mode in the camera app—will get better while using fewer computing resources. But it’ll also enable things that weren’t possible before.
His first example is a photo captured by a Google engineer’s child waving at the camera indoors. It was taken with the Pixel 6 Pro, but with all the AI and ML functions disabled. The lighting isn’t great, the child is moving and so is the person holding the camera, which is why the child’s face and hand wave are blurry. Osterloh then compared it to a photo captured with the AI and ML models turned on. The features of the child's face became clearer. “What we want to do is give the user the intent they have and capture this really fun moment,” he says.
To achieve this result, the Pixel takes a photo from the main sensor with a normal exposure, and then another from the ultrawide camera with a fast exposure. It combines the two, so it’s getting the sharpness from the ultrawide but accurate colors and noise from the main sensor. The Tensor chip then corrects for motion—the subject’s hand wave, as well as the camera shake induced by the photographer’s hand. It’s also running a face detection model to identify and ensure the subject’s face is prioritized to be in focus.
“This is done with a series of different sophisticated machine learning models running parallel in real time, fusing images from multiple different sensors at once,” Osterloh says. “That’s what we wanted to do with Tensor, and it’s what makes it possible.” The Pixel camera is smart enough to know that it doesn’t always need to use both sensors for every photo, just when it’s facing a challenging situation.
Another area of improvement is video quality. Pixel phones have been criticized for lackluster video performance compared to their still-photography prowess.
“It’s really hard to run a lot of the same algorithms and machine learning techniques on video because there’s so much processing happening,” Osterloh says. “We were really limited and weren’t able to do any of that transformation in the past, but with Pixel 6 we are. With Tensor, we are.”
Osterloh shows off a video comparison taken by three phones: the iPhone 12 Pro Max, the Pixel 6 Pro, and the Pixel 5. The subject is a field by a lake during sunset, a challenging scene that will test any phone’s HDR capabilities. The Pixel 5’s video has an over-brightened sun, and the grass is too green, not the more realistic golden yellow hue you’d expect from a sunset. The iPhone 12 Pro Max’s clip is closer to what you’d see in real life, but it’s washed out in some areas, and the white balance of a nearby tent is off. The Pixel 6 Pro’s footage offers up true-to-life shadows while maintaining natural colors you’d expect to see in such a scene.
This was a controlled test, so it’s difficult to say exactly how much better the video quality will be on these new Pixels. For example, Osterloh didn’t have any clips to show off the Pixel’s video stabilization, which iPhones excel in. Still, in the brief demo, the Pixel 6 appeared to capture crisper video, with more accurate colors, than the video captured on Pixel 5.
Aside from improving camera performance, the Tensor chip affects every other part of the Pixel. There will be security improvements, Google says, with an updated version of the company's Titan M security coprocessor running alongside the Tensor chip. (The company says it plans to release white papers with more security-related details around the time of the Pixel's launch.)
Two other Tensor-related updates Osterloh showed off are related to Live Transcribe, which automatically transcribes videos on the fly, as well as voice dictation in Google’s keyboard app, Gboard.
Transcribing a video in real time on device typically requires intense computing power. With Tensor, Pixel 6 phones are able to transcribe and translate simultaneously, as evidenced by a video of a French presenter whose words the Pixel converted to English at the bottom of the screen. (Whether the translations are accurate, I cannot say.) This also has battery life implications: Osterloh says this was running at about 50 percent power consumption as the Pixel 5.
With Gboard’s voice dictation, Osterloh talked to the Pixel to write and send a message. The dictation was rapid, and while it initially missed one punctuation mark, it added it back in after Osterloh finished speaking because it understood the context of the entire paragraph. You can also say “Clear,” or “Send” to erase or send a message; the software will understand your tone and intent, so it’ll know when those are commands and not words that are a part of a sentence.
The Pixel 6 and Pixel 6 Pro
Google's big reveal of its Tensor-powered software features is a departure from typical smartphone launches: Usually, it's the shiny new hardware that's shown off first, followed by a deeper dive into the phone's functions. In this case, the Pixel 6 hardware specs aren't being revealed just yet.
In briefings with WIRED, Osterloh had both the Pixel 6 and Pixel 6 Pro on hand to show off all of these improvements. It was obvious that the two new phones share the same ultrawide camera and main sensor, which can gather 150 percent more light than the Pixel 5, enabling brighter and sharper low-light images. The Pro is the bigger phone, though, and it has an additional 4x optical telephoto camera lens.
However, the phone's external design has been completely revamped. Instead of a standard left-aligned camera bump, all the cameras sit in a thick horizontal bar that spans the entire back of the phone. There are an array of colors to choose from, and the top part above the camera bar has an accented color, mimicking the two-tone design of the original Pixel.
“If we’re going to transform what’s under the hood so much, we should really transform the design of the product too,” Osterloh says. “Pixel has always been known for photography, and this just makes that point very clear.”
Aside from the center-mounted selfie camera and an in-display fingerprint scanner, Google isn’t sharing much else about the Pixel. Osterloh says the company will go into greater detail in the fall, closer to the Pixel 6's official launch. He also didn't share pricing for the new phones, or for the upcoming Pixel 5A—which is expected to be cheaper than a flagship model.
Not planning on upgrading, either way? Osterloh says Google still plans to bring some new features to older Pixels, as it has done for the past few years, but it’s still dependent on whether the hardware can handle it. In more positive news, it’s likely the Pixel 6 line will see longer software support. Competitors like Samsung are delivering four years of security updates and three Android OS upgrades, yet Google has been stuck on three years of support.
“We’re working on what our final numbers will be there and we definitely see a possibility to extend it, so we’re excited about that,” Osterloh says.
Google's investment in its own chips underscores the company's commitment to consumer hardware, even though Google still only claims a single-digit percentage share of the global smartphone market. Osterloh indicated that the Tensor chip might show up in other consumer devices over time, which would help it amortize the cost of its investment in TPUs.
“Building your own chip can only last as long as you can scale, and it requires a lot of people hours to maintain that in a competitive landscape,” says Anshel Sag, a senior analyst at Moor Insights & Strategy. “That's why Apple built its chips across iPhone, iPad, and Mac, and amortize those chips across so many models."
So how will Tensor compare to Apple's Bionic chips, which the Cupertino company likes to tout as performance powerhouses? Osterloh wasn't ready to share benchmarks on the new Google chips. Sag says he believes it would be difficult for Google to accomplish something in mobile chips that hasn't already been done. “I don't see Google out-engineering Apple's team right off the bat, because Apple has had multiple generations of success in chips," Sag says.
One element of Google's smartphone business that could work in its favor, though, is the relatively low volume of phones it sells. The global squeeze on semiconductors is expected to last for many more months, and is affecting everything from auto shipments to home appliances to gadgets. If Google is shipping this system-on-a-chip this year, Sag says, the company likely ordered wafers around two years ago and have almost certainly negotiated for exactly how many phones they expect to sell. “Which is not many,” Sag concludes.
“Pretty much everything in semiconductors is constrained right now," Osterloh says. “For this particular chip, we have a lot of control over it, and we don't believe the Pixel 6 line will be constrained. So that's a positive. But there are so many semiconductors in the product, like in every electronics product, that it's definitely been a difficult problem. It's affected us this year, for sure.”