Google said that it plans to use its own in-house processor to power its upcoming line of Pixel smartphones when it ships in late 2021.
The technology giant said the chip, called Tensor, was specifically designed to assist with artificial intelligence workloads in its Pixel 6 and Pixel 6 Pro phones. Google said the chip would give it the computing resources it needs to roll out more advanced features in areas such as image processing and speech recognition it hopes will help differentiate its devices from Apple’s iPhone and Samsung’s Galaxy, which both use bespoke silicon.
Google CEO Sundar Pichai called the custom-designed part its “biggest innovation in Pixel to date.”
The chip takes over the slot held by Qualcomm’s Snapdragon chips in every generation of Google’s Pixel, which the technology giant rolled out in 2016. The current Pixel 5 is based on the Snapdragon 765G chip
For years, Google-developed devices put software and AI features over hardware. But with its internally designed chip, the company hopes that it can more tightly integrate the software and hardware at the heart of the Pixel 6 and Pixel 6 Pro phones, reducing power and giving it a performance hike in the process. Bringing chip development in-house could also give it more control over its future roadmap and component costs.
The chip is a reference to the Tensor Processing Unit (TPU) used by Google in its data centers to run machine learning chores. The chip will probably have a dedicated machine-learning module, potentially a smaller version of the TPU. The module could be used to run machine learning (also called inferencing) directly on the device without offloading to vast warehouses of servers in the cloud to process all the data. It is generally more secure to run inferencing on a device and it also assists in reducing delays due to latency.
“Tensor was built for how people use their phones today and how people will use them in the future,” said Rick Osterloh, Google’s senior vice president of hardware and service, in a blog previewing its upcoming Pixel.
Google is not sharing any further details about the upcoming Pixels or performance metrics of the Tensor chip. But it outlined some of the philosophy behind its new Tensor chip as part of the preview announcement.
Tensor is a system-on-chip (SoC) that integrates all the Pixel’s major processing components on a single die. The new processor will probably have roughly the same lineup of building blocks as the Snapdragon chip in its current-generation Pixel 5, which contains a central processing unit (CPU), graphics processing unit (GPU), image signal processor (ISP), a machine-learning engine and a range of other intellectual property (IP) blocks.
The CPU is likely based on the Arm architecture. Both Apple’s and Qualcomm’s chips, which are used in hundreds of millions of smartphones globally, leverage Arm’s technology to build the central processors. But it is unclear whether Google designed the CPU core from scratch or used ready-made blueprints from Arm. It is also unclear what technology node Tensor is based on and what made-to-order foundry is manufacturing it.
The devices also will have the latest generation of the Titan M chip—an ultra-secure processor internally developed by Google that it first added to the Pixel in 2018.
The Titan M1 is used in the latest Pixel smartphones to run checks every time the system starts to make sure nothing has been maliciously altered. If anything is wrong, it blocks the phone from booting up, preventing attackers from altering the firmware in the phone to steal information. The chip can also prevent attackers from returning Android to a previous, potentially vulnerable model of the operating system (OS), and block anyone trying to unlock the bootloader and gain access to more critical levels of the software of the device.
The co-processor can also double as a secure vault for passwords as well as secret keys used in encryption.
The Titan M2 will probably have an expanded set of security features. A secure core housed in the Tensor chip will work with the second-generation Titan M chip to add another layer of hardware security to the Pixel.
Designing a chip is a risky bet in the consumer electronics business, where shorter and shorter product cycles can often make even the most advanced processor obsolete—or at least outdated—in less than a year. But the pursuit has become trickier and more expensive than ever in recent years. If Google wants to warrant the hundreds of millions of dollars that can go into a custom processor design, the Tensor chip will require a major leap in performance over merchant chips from the likes of Qualcomm or MediaTek.
Based on the 5-nm technology node, Qualcomm’s flagship Snapdragon 888 brings a 25% boost in performance per watt to the CPU inside while supplying 26 trillion machine learning operations every second.
Google said that it also designed the Tensor chip to run its computational photography models on the devices, delivering new features to users plus improvements to existing ones: for instance, its Night Sight mode that can shoot subjects in very dark conditions without a flash. Google’s Super Res Zoom technology can improve zoomed-in, low-resolution images by snapping several shots and then extracting extra detail.
Google said the Tensor chip will also improve the voice-processing features in its devices. Google introduced the Recorder function with its Pixel 4 smartphone, which uses machine learning to record and then transcribe speech to text, complete with punctuation, in real-time—without any connection to the cloud. Google has also improved the ability of its Google Assistant to listen to speech and interpret voice commands all on its Pixel 5.
“As more and more features are powered by AI and ML it’s not simply about adding more computing resources, it’s about using that ML to unlock specific experiences for our Pixel users,” said Osterloh.
The chip also opened the door for Google to overhaul the camera system in its new Pixel phones. The improved image sensors and lenses in the Pixel 6 and Pixel 6 Pro are too large to fit into a traditional square-shaped module. Google said that it upgraded the “Pro” model with three cameras arranged in a “bar” on the back of the phone. The Pixel 6 will ship with dual cameras, also packaged in a rear-mounted bar, Google said.
The chip highlights how Silicon Valley giants are building up in-house semiconductor expertise to design chips that better suit the requirements of their products and give them more control over their futures.
Google is stealing a page from the playbook of its larger rivals in the smartphone market. That includes Apple, which has designed the A-series SoCs in its iPhones for more than a decade to reduce dependence on suppliers, limit costs, and integrate hardware and software in a way that differentiates it from rivals. Now Google is giving itself more control over the silicon (Tensor), the OS (Android), and system hardware (Pixel) in the same way Apple controls the same three building blocks with its A-series processors, iOS, and iPhone.
Google has a history of in-house semiconductor development. In addition to the TPUs running machine learning in its cloud data centers and the Titan M co-processor in its current family of flagship phones, the company internally designed a server processor called Argos to speed up video transcoding and improve the transmission of high-quality video to users. Reportedly, thousands of the chips are running in its data centers.
Google has also rolled out a tiny AI coprocessor called the Edge TPU. The chip is part of its Coral product line and performs 4 trillion operations per second (TOPS) at 2 W in a package that measures only 5-mm × 5-mm.
Google is not one of the leading players in the smartphone market, where global shipments improved by more than 13% year over year to more than 310 million in the second quarter of 2021. Samsung was the world’s largest smartphone vendor in the quarter, finishing ahead of Apple, Xiaomi, and others in terms of market share. Industry analysts estimate that Google’s share of the global smartphone market stands at around 1%.
That could prevent Google from reaping a major advantage to in-sourced chip development: the cost savings. “Without scale, building a custom SoC is a losing proposition, and if a design team isn’t competitive with what’s already available in the market today, it comes off as a loss-leading exercise,” said Anshel Sag, industry analyst at market research firm Moor Insights and Strategy, in Forbes. “There is an established track record of this being done by smartphone manufacturers, but generally ones with much higher volumes than Google.”
But he adds that it is also possible the company moves Tensor to other product lines in the future. “If Google builds a powerful enough SoC, it could foreseeably use these same SoCs in its tablets and Chromebooks.”