A Turning Point for Machine Learning: Google Unleashes First In-house Tensor SoC

October 23, 2021 by Jake Hertz

Joining the ranks of Apple with releasing its first custom in-house SoC comes Google. With the release of Pixel 6, Google makes a splash in the machine learning realm with its Google Tensor.

Google is often at the forefront of software but is essentially not considered a hardware company. While this is mostly true, one place where Google is, in fact, a hardware company is due to machine learning (ML). 

Most notably, Google's history with ML hardware includes their Tensor Processing Unit (TPU), an AI accelerator that comes as an edge device or a cloud device. 

This week, Google has expanded its hardware portfolio with the release of Google Tensor, its first-ever in-house custom system on chip (SoC), designed for the Pixel 6


Google's first in-house custom SoC.

Google's first in-house custom SoC. Image used courtesy of Google


In this article, let's look at the new release from Google to see what Tensor has to offer.


Tensor's Computing 

Tensor is an ML-focused SoC meant for smartphones, and as such, aims to integrate several powerful hardware resources into a small form factor.

Tensor's heart is the CPU architecture highlighted by two high-performance ARM Cortex-X1 cores, both topping out at 2.8 GHz. 

The high-performance cores, which handle the heavy processing, are complemented by two 2.25 GHz A76 cores followed by four high-efficiency A55 cores. This eight-core architecture is said to allow Google to balance performance and efficiency and avoid maximum utilization of one individual resource. 


A system-level block diagram of the Google Tensor.

A system-level block diagram of the Google Tensor. Image used courtesy of Google and 9TO5Google


Being an SoC, Tensor features more than just the CPU architecture. For graphics and gaming tasks, Tensor integrates a 20-core, Mali-based GPU. 

Further, Tensor features a TPU for ML processing, an image signal processor (ISP) built specifically to allow efficient 4k HDR+ video, and a Context Hub meant to enable always-on machine learning tasks at ultra-low power. 

Finally, Google hones in on security with Tensor by integrating a Titan M2 security core on-chip. Combined with Arm's TrustZone security technology, the Titan M2 security processor hopes to protect sensitive user data on the Pixel 6. 


Tensor's Performance 

In terms of compute, Google claims [video] that Pixel 6's Tensor hardware outperforms Pixel 5's hardware tremendously. The Tensor CPU is said to be 80% faster than Pixel 5's Snapdragon 765G. On the other hand, the Tensor GPU is said to be 270% faster than Pixel 5's Adreno 620 GPU. 


CPU and GPU performance of Pixel 6 vs. Pixel 5.

CPU and GPU performance of Pixel 6 vs. Pixel 5. Screenshot used courtesy of Google [video]


In terms of power, many applications, such as face recognition and voice assistant, are said to operate at 50% of the power consumption of Pixel 5. 

These improvements are primarily due to the advanced hardware and architecture used in Tensor but are also because of improvements in software. 

Google also gives credit to advances such as the Android 12 OS, a brand new neural machine translation (NMT) model, and a new speech recognition model, which Google claims is the most advanced they've ever released. 


A Custom In-house Chip Fabrication Future 

With Tensor, Google hopes to make Pixel 6 the most advanced and capable smartphone on the market while maintaining significant battery life. 

Also, with this release, Google joins the likes of Apple and Samsung in making their own smartphone silicon. 

As Pixel is one of the biggest competitors to Samsung's Galaxy and Apple's iPhone, this move may represent an exciting industry trend, where big-time smartphone makers are finding more value in creating their own silicon than using outside chips like the Snapdragon. 

It will be interesting to watch how other major players, like Huawei, follow suit and what this could mean for the industry.