Tensor is Google's most important launch of 2021, not the Pixel 6

This chip represents Google's plan for the future

The Google Pixel 6 and Pixel 6 Pro are indisputably landmark phones that we’ll be talking about for years to come. They bear a bold and immediately recognizable style, and are setting a new bar with their combination of cameras and post-processing capabilities. While these qualities are great for attracting customers, Google likely sees its first custom-designed Tensor chip to be the most important part of this launch. Not only does it factor into ambitious features like live translation and ambient computing, but this chip also signals a strategic shift for Google as a hardware manufacturer, and perhaps hints at greater plans for the future.

But what is Tensor? To write it off as another system-on-a-chip (SoC) would be fair in casual conversation, but there’s much more to it. Of course, Google isn’t throwing open the doors to all of the intricate details, but thanks to the company’s marketing and blog posts, leaks and analysis by community members, and interviews held with Ars Technica and CNET, we can piece together quite a few details about the components and how they are utilized.

Specs

Let’s begin by looking at the main components. Like most other mobile chips, Tensor is made up of an array of physical cores, some of which have been designed to run efficiently on low power while performing background tasks and maintaining idle states, while the other cores handle larger tasks very quickly at the cost of more power and heat.

For example, this year’s de facto flagship chip for smartphones is the Qualcomm’s Snapdragon 888. It features a single Cortex-X1 “supercore” for big tasks, three Cortex-A78 “medium” cores, and four low-power Cortex-A55 cores. The X1 will run for short bursts to handle big jobs, but it will be shut down quickly to reduce power consumption and heat. The medium cores are there for sustained workloads and the majority of standard apps; but again, Android attempts to limit their use.

Google worked with Samsung to build Tensor using Exynos SoCs as a template. That happens to involve using the same cores built on Samsung’s 5nm LPE manufacturing process, but Google broke from the traditional design to use an uncommon configuration. The four A55 cores remained for background tasks, but Google opted to double up on the big X1 cores. And instead of using three A78 cores, which are the “medium” cores of this generation, Google slotted in two A76 cores that had previously been used as "big" cores.

According to Phil Carmack, vice president and general manager of the Google Silicon team, this arrangement allows Tensor to split demanding workloads between the pair of high-power cores. Since they can finish tasks more quickly without producing as much heat, they’re actually more efficient than running a single core for longer.

As it turns out, the X1s will also be used for some moderate tasks that would normally have gone to the A76 cores. Carmack says that the X1s barely have to work for many of those jobs, which allows them to be run at low frequencies, but the same workload would max out an A76. This may mean the A76 cores are just there to handle chores requiring a little more speed than an A55 can deliver, but too small to justify spinning up an X1.

Google says Tensor’s CPU performance is 80% faster than the Snapdragon 765G found in last year’s Pixel 5, and the GPU performance is up to 370% faster thanks to the 20-core ARM Mali-G78 MP20. If we’re being honest, these claims are somewhat undermined by the fact Tensor is being compared to an older chip that was only intended for mid-range phones. Nevertheless, reviews of the Pixel 6 confirm the device is very responsive and snappy where it needs to be; and despite some mixed results in benchmarks — especially for gaming — it’s reasonably comparable to the Snapdragon 888 and Exynos 2100. For a deeper dive and analysis on the architecture, check out AnandTech’s article.

Smart bits

Top-tier specs and some subtle improvements to user experience are nice, but these hardly warrant the 4 years Google says it has spent developing a custom chip if all it can do is tie the competition. The true benefits of Tensor come from Google’s custom cores.

Monika Gupta, senior director of the Google Silicon team, says Tensor is designed with a strong focus on artificial intelligence (AI) and machine learning (ML). In fact, the chip’s name is inspired by the Tensor Processing Units (TPU) that make up part of Google’s Cloud infrastructure, which is itself designed for the TensorFlow software platform.

Of course, Google isn’t new to building custom cores. These efforts began with the Pixel Visual Core, and later the Pixel Neural Core. Google also developed the Titan M security chip, a full coprocessor that’s physically distinct from the SoC.

So what is actually included in Tensor beyond CPU and GPU cores? The usual suspects are present, like video encoder/decoders, audio, memory, and so on. However, there are also a few custom components that play a big role in Google’s vision of the Pixel experience.

First and foremost is the Tensor Processing Unit (TPU), more conventionally known as a Neural Processing Unit (NPU). This is where the majority of AI and ML operations will occur, and it’s responsible for most of the Pixel 6-exclusive features like automatic speech recognition for live captions, live language translation, and computational photography (e.g. Motion Mode, Action Pan, Long Exposure).

Some members of the modding community have successfully activated Pixel 6 exclusive features to run on other devices, and they usually work, but not as quickly, and not without heating up and rapidly running down the battery. This lends some credence to Google’s claims that Tensor is necessary to accomplish some of these ambitious features — not because it’s impossible on other chips, but because they’re too demanding to run well without optimized hardware.

Tensor also includes an enhanced Image Signal Processor (ISP) chain. This is the portion of an SoC responsible for interfacing with the cameras, turning raw data into pixels, and making corrections like reducing noise and correcting colors. Google has credited two particular features to the new ISP: Real Tone for accurately representing skin tones, and HDRNet for real-time processing of video with the same techniques that were previously exclusive to stills. Rick Osterloh told the Verge that the ISP is directly linked with the TPU, effectively lending a lot of extra horsepower to the processing requirements.

On the other end of the spectrum is the Context Hub, which Google frames as machine learning for low power operations. This is where Tensor will run tasks that have lower priority, especially during idle states, and they’ll have less impact on your battery. Or to put it simply, Context Hub is the champion for ambient computing features like Now Playing.

Security has also been a regular talking point for Google, and a lot of attention has been put on the new Titan M2 coprocessor. The Titan M2 happens to be physically separate from the Tensor chip, but it’s tightly coupled with the dedicated Tensor Security Core inside of the SoC. This core is running Google’s open-source Trusty OS and maintains a secure environment from the moment a Pixel 6 boots. The Security Core and Titan M2 are responsible for storing and working with user data keys, cryptographic operations, and many of the other authentication and verification tasks that are called for.

Why build it?

Tensor development began in 2017, around the same time Google announced plans to acquire HTC’s Pixel team. In collaboration with Google Research, the Silicon team set out to design a chip based on the future of machine learning, and to put some of that technology to use for new experiences that would give Pixel phones room to do bigger things.

Gupta explained that this came about after seeing so much technology inside of Google that couldn’t be implemented on Pixels with the SoCs available on the market.

Looking toward Apple gives us another reason Google may have chosen this path. It’s no secret that Apple’s switch to custom silicon in the iPhone line has been good for both boosting performance and cutting component costs. The iPhone transition proved so successful that it led Apple to make the same move with its Mac lineup.

Tensor certainly does achieve some big performance gains over currently available chips, but they’re very narrowly focused on machine learning operations. This brings plenty of benefit to Google’s own applications, and some limited benefit to outside developers with ML-oriented software. However, unlike Apple’s A-series chips, Tensor doesn’t really improve anything for the larger scope of apps and games. But it’s fair to remember that this is just the first generation to see daylight, so it’s possible Google will continue to expand its efforts to more components in the future.

As for cost, we’ll probably never know if Google is saving money with Tensor, especially if the market share for Pixel phones remains low compared to its leading competitors. But the potential also isn’t exclusive to phones, Google stands to see greater returns on this investment as Tensor is adapted for future Chrome OS and Wear OS devices.

But this leaves us with a bigger question: What comes next for Tensor? Google’s new chip poses a threat to several different competitors, particularly chip makers like Qualcomm, but also other phone manufacturers like OnePlus.

Google is now in a position to expand its partnership with Samsung and begin selling future generations of Tensor chips (or at least the custom cores) to other OEMs. Not only is this a profitable avenue, it also expands the reach of Google’s most advanced ML-based software and services to many non-Pixel devices. Of course, that may also involve going into direct competition with Qualcomm. On the other hand, Google may choose to keep Tensor in-house to avoid a conflict of interest, but this inhibits the software advantages and may also encourage OEMs to align more closely with other chip manufacturers to create their own competing solutions.

Google’s new chip is almost certainly more than just a way to add a few clever software features to the company’s latest smartphone. Whether this turns out to be another short-lived Google experiment or becomes a whole new side of the business, Tensor has the potential to kickstart the next chapter for smartphones.

About The Author