Google’s Pixel 6 is shaping up to be the company’s most ambitious smartphone in years, largely due to the new, custom-designed Tensor processor, which aims to catapult Google to the forefront of the smartphone market with the power of Google’s years of machine learning experience. And Google needs it: despite the popularity of Pixels in tech circles, its phones just aren’t popular sellers in the US, barely moving the needle compared to juggernauts like Samsung and Apple or even smaller players like Oppo or Xiaomi.
Tensor is Google’s big bet, centered on the AI-boosting TPU that promises to improve photos and videos, search, captioning, text-to-speech, and more. It’s a tall order for any chip, much less one focused mainly on machine learning as its standout feature — but while the Tensor SoC might not launch Google to iPhone-like heights just yet, it could be a crucial first step toward turning its Android also-ran into a top contender.
The rest of the SoC is a mystery right now, but it seems like Google will be using third-party designs for things like the CPU, GPU, and modem — meaning that the Pixel 6 will probably feel pretty similar to any other Android smartphone powered by a Qualcomm or Samsung processor for most tasks, instead of some sort of revolutionary upgrade on par with Apple’s A-series powered iPhones.
Google hasn’t given a lot of information about what Tensor’s actual architecture is going to look like for things like the CPU, GPU, modem, or other major components of the SoC beyond the TPU. But based on rumors and the fact that Google isn’t taking the chance to crow about any major customization or advances its made here, it’s likely the case that most of Tensor’s hardware stack will be outsourced designs. Qualcomm and Samsung already do something similar — the Snapdragon 888 uses partially customized versions of Arm’s Cortex-X1, A78, and A55 designs, while Samsung’s Exynos 2100 uses Arm designs for both its CPU and GPU.
XDA notes more specifically that the Tensor will likely be some combination of Arm’s Cortex-A78, Cortex-A76, and Cortex-A55 CPU cores and Arm’s standard Mali GPU. Which means that the difference between Tensor and, say, a Snapdragon 888 or Exynos 2100 might not be that major for things like overall CPU or GPU performance. That’s a good category to be in, especially if Google is trying to truly make a flagship device.
But as Google’s Rick Osterloh told The Verge, “the standard stuff people look at will be very competitive and the AI stuff will be totally differentiated.” The things that make Tensor special and unique aren’t how fast it can run games or how efficient its CPU and battery life is, though.
That means that Tensor probably also won’t be the magical solution that Android fans have been hoping for: a custom, Google-made chip designed specifically for Android and the Pixel’s hardware to deliver the kind of performance and power that Apple’s been able to offer on the iPhone for years using a similar strategy.
Tensor is just a first-generation product
But the good news is that Tensor is just a first-generation product; it’s easy to point at Apple’s custom chips and demand that Google should do the same, but it’s important to remember that Apple’s earliest A-series iPhone chips started in a similar fashion. The A4 and A5 chips featured standard Arm designs for the CPU cores (with some optimization and improvements on top), before Apple switched to fully custom designs in later generations.
Google is rumored to be working with Samsung, which will both be manufacturing the chips — which makes sense, given that it’s one of the only two companies, along with TSMC, currently fabricating 5nm chips (a service it provides for numerous companies, including Qualcomm).
Samsung is also rumored to be working more closely with Google on the actual design of the chips, too, leveraging its Exynos hardware and software in a capacity beyond simply manufacturing. We’ll have to wait until Google reveals more about Tensor to know the extent of that partnership, but it could mean that Google will get some extra help in more substantial custom chip improvements than if it had to reinvent the wheel from scratch.
The idea of a purpose-built machine learning chip in a Pixel phone isn’t a new idea for Google either. The company has put dedicated in-house AI chips in its phones before, like the Pixel Neural Core and Pixel Visual Core in the Pixel 2, PIxel 3, and Pixel 4 models, making the Tensor less of a brand-new advancement and more a refinement of its past work.
Tensor’s TPU does seem to be the next step from that, presumably offering more powerful AI functionality than either of those two previous chips. It’s also more closely integrated with the phone, with Google noting that it’s able to do things like directly route image data through the TPU.
Those benefits may be worth the development of a custom SoC — again, we’ll have to wait and see how Google actually implements those AI features, and how substantial of an improvement they are over a standard Android flagship. (A sentiment that, in general, applies to how heavily Google is emphasizing the TPU as the major differentiator in the Pixel 6, something that feels like a tricky thing to actually sell customers on.)
As it stands now, though, the main difference between Tensor and a Snapdragon 888 looks like it’ll just be some fancy neural network tricks. But, much like with Apple’s early in-house chips, Tensor may just be the first step in a longer journey toward a more bespoke Google chip. And that’s a far more exciting concept than subtly better machine learning demos.