Google is announcing the Pixel 6 and Pixel 6 Pro today, though it might be better to call it a preview or a tease. Rather than releasing all the details on its new Android phones, Google is instead putting the focus on the new system on a chip (SoC) that will be inside the new Pixels. It’s called the Tensor SoC, named after the Tensor Processing Units (TPU) Google uses in its data centers.
Tensor is an SoC, not a single processor. And so while it’s fair to call it Google-designed, it’s also still unclear which components are Google-made and which are licensed from others. Two things are definitely coming from Google: a mobile TPU for AI operations and a new Titan M2 chip for security. The rest, including the CPU, GPU, and 5G modem, are all still a mystery.
Less mysterious: the phones themselves. I spent about an hour at Google’s Mountain View campus last week looking at the phone hardware and talking with Google’s hardware chief Rick Osterloh about Tensor. After all that, my main takeaway about the new Pixel 6 phones is simple.
Google is actually, finally trying to make a competitive flagship phone.
This fall, Google will release two slightly different Pixel phones: the Pixel 6 and the Pixel 6 Pro. If the final versions are anything like the prototypes I saw last week, they will be the first Pixel phones that don’t feel like they’re sandbagging when it comes to build quality. “We knew we didn’t have what it took to be in the ultra high end [in the past],” Osterloh admits. “And this is the first time where we feel like we really have it.”
Both versions of the Pixel were glass sandwiches with fit-and-finish that are finally in the same league as what Samsung, Huawei, and Apple have to offer. “We’ve definitively not been in the flagship tier for the past couple years, this will be different,” says Osterloh. He also admits that “it will certainly be a premium-priced product,” which I take to mean north of $1,000.
Google is only sharing a few of the key specs for each phone, leaving the details for later — likely October. (And no, there was no mention of a folding phone nor a watch.) Google also wouldn’t allow us to take photos or video of the devices during our meeting. In any case, here is what we do know:
The Pixel 6 Pro will have a 6.7-inch QHD+ display with a 120Hz refresh rate. That screen is very slightly curved at the edges, blending into shiny, polished aluminum rails on the side. It has three cameras on the back: a new wide-angle main sensor, an ultrawide, and a 4X optical-zoom folded telephoto lens. Google isn’t sharing specs on the camera beyond saying the main wide-angle sensor takes in 150 percent more light.
The regular Pixel 6 has a 6.4-inch FHD+ screen with a 90Hz refresh rate. Its screen is perfectly flat, with matte-finished rails. It also loses the telephoto camera.
Although there will be memory differences between the phones, both will have the new Tensor SoC, a Titan M2 security chip, and in-display fingerprint sensor. There will be slightly different color options for the two types of phones.
As is often the case with polarizing designs, the look of the Pixel 6 and 6 Pro make a little more sense in person than in leaked images. There is a huge “camera bar” that runs the full width of the phones, with a barely raised metal rail to protect the glass from scratches. There are only so many ways to handle massive camera bumps on big phones and Google’s solution is to “celebrate and highlight” them, in Osterloh’s words.
If all Google were doing was rescuing the Pixel line from the doldrums of the midrange, that would be significant but not worth a months-early pre-announcement. The most important part of the Pixel 6 and 6 Pro is that Google is using that new Tensor SoC inside.
So excited to share our new custom Google Tensor chip, which has been 4 yrs in the making ( for scale)! Tensor builds off of our 2 decades of computing experience and it’s our biggest innovation in Pixel to date. Will be on Pixel 6 + Pixel 6 Pro in fall. https://t.co/N95X6gFxLf pic.twitter.com/wHiEJRHJwy
— Sundar Pichai (@sundarpichai) August 2, 2021
Qualcomm has a virtual monopoly on processors in Android devices in the US. Worldwide, there is a little bit more competition as Samsung, MediaTek, and Huawei all have chips in Android phones. But on the whole, processing power on Android phones is rightfully thought of as woefully behind what Apple has done with its own in-house silicon on its A-Series line of chips.
Because of that situation, there’s a lot of interest to see if Google could potentially make a more competitive chip that could differentiate its products. But don’t let that interest trick you into thinking that Tensor is exactly equivalent to Apple’s A-Series chips. Tensor is the system on a chip, with a mix of components that Google itself has designed and others that it has licensed.
Google’s not sharing who designed the CPU and GPU, nor is it sharing benchmarks on their performance — though Osterloh says that it should be “market leading.” (Current rumors suggest that it might be Samsung providing those more standard component designs.) He adds, “The standard stuff people look at will be very competitive and the AI stuff will be totally differentiated.”
Instead, this week’s announcement is an attempt to reframe the narrative away from gigahertz and toward artificial intelligence and machine learning in phones — areas where Google, of course, has a big advantage.
Typically when you think about a phone’s specs, you think of the core three: CPU, GPU, and RAM. Those pieces of the SoC are what impact your day-to-day experience the most — how fast the phone feels, how long it lasts on battery, how well it connects to a cellular network, and so on. After that, there are generally some co-processors off to the side that handle discrete tasks like image processing or security. Google itself has already made some of those — the Titan M chip and Pixel Visual Core have appeared on previous phones.
“It’s definitely very different than just another co-processor,” Osterloh says. “Like with any SoC, we license a lot of technology into it, but this is our design and it was designed specifically with the purpose of driving our ML and AI forward.” Google’s argument is that the new chips in Tensor are an essential part of many of the things the new Pixel phones can do — not unlike Apple’s Neural Core in its A-Series processors.
“The computers of the future are becoming much more heterogenous than they have in the past,” Osterloh argues. He has for years been signaling that the end of Moore’s law will mean that computers — and phones — will need to be built differently. “There’ll be a lot more specialized sub-elements to the design to be able to do things in specific ways. This is a consequence of raw computing power running out of headroom or growing more slowly than the kinds of processing we want to do with AI,” he says.
The most important of those chips is a mobile version of a Tensor Processing Unit. Google has been making TPUs for its server farms for over five years now, dedicated to more efficiently performing AI and ML tasks. It offered an “edge” version of its TPU for enterprise solutions a few years ago, but the Pixel 6 marks the first time Google has put a mobile TPU in a phone.
So what can the new TPU inside Tensor actually do? Google had a small handful of demos. The first two were, unsurprisingly, related to photography. Using ML to make better photos has historically been a huge advantage for Pixel phones, but in recent years progress has stagnated and competitors have caught and surpassed the Pixel — and far surpassed it when it comes to video.
Google clearly wants to take back the crown and thinks the TPU is the way to do it. The first demo Google showed was a blurry photo of a toddler — the kid was moving because that is what kids do. A second version of the photo was the same but run through Tensor’s TPU, and the kid’s face was sharper.
It wasn’t a deepfake. What the Pixel 6 does is run the process of snapping a bunch of photos from the main sensor and combine them into a single, HDR image. But now the Pixel 6 also has the ultrawide grab a fast, sharper image to capture that detail. Then the TPU recognizes that there’s a face, and then combines the ultrawide’s image with the rest.
But impressive image demos are a dime a dozen these days. Google itself famously promised to remove chain link fences from photos in 2017 but never delivered. The promise of Tensor is to deliver better results more quickly, since the SoC is designed to more efficiently run data through the TPU.
It’s right to be skeptical until we get a chance to test the camera ourselves. It’s necessary to be skeptical of the Pixel 6’s video abilities given the Pixel line’s history of mediocre video — but Google’s second demo didn’t make skepticism easy.
It was a simple pan across a beach, with the setting sun fully in frame for much of the shot. As an HDR video, it was challenging. Google set up a rig with the Pixel 6, Pixel 5, and iPhone 12 Pro Max and shot the same pan with all three. As you might expect in a demo provided by Google, video from Google’s new phone looked the best.
Specifically, it didn’t artificially brighten shadows too much like the iPhone 12 Pro Max and also maintained a more natural white balance throughout. Compared to the Pixel 5, it was no contest. Last year’s Pixel over-sharpened everything into abstract art while the Pixel 6 looked much more natural.
The main reason the Pixel 6’s video was better, according to Osterloh, is that putting the TPU in line with the whole image processing stand means that the same HDRNet process that Google applies to still images can now be applied to every single frame in video. The demo I saw was in 4K at 30fps.
The other demos were a bit more subtle and were related to translating speech to text, which Tensor handled locally without needing an internet connection. In one, the Pixel’s on-device auto-caption feature appeared to be a little faster and more accurate — and was even able to translate from French to English in real time from a playing video. “We’re now able to run data-center quality models on our device,” says Osterloh.
In another demo, Osterloh showed that speaking to type was nearly instantaneous and that he could use the keyboard to edit words inline at the same time he was speaking — both input methods were active at the same time.
In addition to the TPU, the Pixel 6 will also have a new version of Google’s Titan M security chip. In the blog post announcing the Pixel 6, Google is going so far as to say that “with Tensor’s new security core and Titan M2, Pixel 6 will have the most layers of hardware security in any phone,” with a footnote that the claim is “based on a count of independent hardware security subsystems and components.”
Finally, Osterloh says there will be an “always-on computer” that will handle low-level, low-power processes like the ambient display. The battery life target for the Pixel 6 is still only “all day,” however.
The thing about AI and ML is that it might make search more accurate and photos better, but it’s not necessarily going to improve everything that happens on your phone. Osterloh suggests that as the TPU takes up more AI cycles, that could free up the other chips for more performance. That may be, but it’s still hard to make AI a selling point for a phone.
So Google has its work cut out for it. Doubly so, actually, since the Google Pixel line has languished in the low single digits of market share in the US ever since it was announced.
With the Google Pixel 6, Osterloh says that’s going to change. He’s ready to start grabbing market share wherever he can get it — whether that be from Apple or Samsung. “The product is really, now, The Google Phone,” Osterloh says. “So we are ready to invest a lot in marketing and we want to grow.”
In last week’s earnings call, Google CFO Ruth Porat warned investors to keep a look out for higher than usual marketing costs: “In addition, we expect sales and marketing expenses to be more heavily weighted to the back half of the year in part to support product launches in the holiday season.”
It has been five years since Google started making Pixel phones and in nearly every single one of those years I’ve asked some variation of the same question: is Google really serious about hardware? Does it intend to be a real competitor and make real money or is the entire effort just a rehashing of the old Nexus strategy? Or, to use a more recent reference, is the Pixel line just in a kind of “Pantry Mode,” kept alive just in case it’s ever needed for real?
Osterloh says that work on this new Tensor silicon began four years ago. Perhaps not coincidentally, it was about four years ago when Google announced it intended to buy HTC’s phone hardware division. That must have been the year Google got serious about the Pixel line — and it’s only now that we’re beginning to see the fruits of that effort.
Will it be enough? Will the Pixel 6 seriously vie for the enthusiast crowd that wants the very best phone and also grab measurable market share at the same time? We are very far from being able to answer those questions right now.
I am not going to tell you to get hyped up for the Pixel 6. It’s too early and there’s too much we don’t know. But I do think it’s very interesting that Google wants to start building up hype.