Google's first Tensor mobile chip is all about AI tricks, but performance is just a passing mention

Posted early from Aofei Temple
Quantum Bit Report | Public Account QbitAI

Google has released a new phone.

But compared to Apple's explosive performance, Google's new products are much more low-key. However, it is worth noting that the new Pixel 6 and Pixel 6 Pro will be equipped with Google's self-developed mobile phone chip Tensor .

As the first self-developed mobile phone SoC, Google said that some AI experience parts will be more personalized and can also reduce power consumption.

What does Tensor contain?

Google's Tensor is planned as follows:

First of all, the CPU part adopts a 2+2+4 eight-core solution: including two Arm X1 cores, two mid-level Arm A76 cores and four high-efficiency A55 cores. The GPU contains twenty cores.

The Qualcomm Snapdragon 765G CPU carried by the previous generation Pixel 5 includes one A76 large core, one A76 medium core and six A55 small cores, and the GPU is Adreno 620.

Overall, Tensor has become the fastest Pixel chip ever, 80% faster than the previous generation Snapdragon 765G, while Qualcomm Snapdragon 888 has improved CPU performance by about 91% over Snapdragon 765G.

Therefore, in comparison, the CPU performance of Tensor is not much different from that of Snapdragon 888.

The highlight is TPU, which is the tensor processing unit. It is Google's unique part and the source of AI computing power for new phones and new features.

The second is ISP, which is a dedicated image processing core.

The Titan M2 security chip is a chip designed by Google to provide protection for mobile phone security, including protecting lock screen passwords and disk encryption.

Placed next to the Tensor security core to prevent advanced attacks such as electromagnetic analysis, voltage glitches, and laser error injection attacks.

The context hub is a low-power coprocessor that helps complete functions such as lock screen display.

Tensor's overall process uses Samsung's 5nm LPE.

What can Tensor bring

Having said so much, let’s take a look at how the new chip performs in practice and where its powerful AI functions are reflected.

For example, in smart calls, in scenarios like answering or dialing China Unicom or China Mobile customer service, you no longer need to remember which button to press and what to select. Pixel 6 will directly present all the options:

In the real-time translation section, the translation quality has improved by about 18%.

A facial deblurring function has also been added to the camera system this time.

If the Pixel 6 camera viewfinder detects a blurry face, it fires up the second camera and takes two photos at once.

The main camera will take regular exposures for low-noise shots, while the ultra-wide camera will take faster exposures to eliminate motion blur.

Google will then align and merge these photos to present you with a clear picture of the face, helping you become a "black photo killer" at the sports venue~

Not only can it partially deblur, but it also has a motion mode that can actively add some blur effects:

There is also magic erase: no more passers-by stealing the spotlight~

Voice typing is faster and more convenient, three times faster than typing by hand, and can correct your misspelled names using your contact list and previous spellings.

as well as anti-phishing protection and detection for calls, text messages, emails, and links, and more.

The new chip will be installed in Pixel 6 and Pixel 6 Pro. Let's take a look at the comparison of mobile phone parameters with the previous generation of models using Qualcomm chips:

Why Google Launches Self-Developed Mobile Phone Chips

From the product itself, Google's original intention of focusing on AI mobile phone chips is to make many excellent AI and ML models practical. From the release of the first generation of TPU in 2015 to its entry into mobile phones today, it has brought users a more practical experience.

The same is true for image processing cores. In 2017, Google launched the Pixel visual core to enhance mobile phone imaging and accelerate camera HDR+ calculations, processing images smoothly and quickly while reducing energy consumption.

The ISP will also continue Pixel's tradition of high-quality photography.

From a market perspective, developing self-developed mobile phone chips can reduce dependence on Qualcomm.

However, the response from Qualcomm Snapdragon's official Twitter account was quite subtle:

Ahem~ (Google and Samsung are going to do something behind my back)

Of course, for Google's long-term development, the launch of Tensor can also take into account both software and hardware and better grasp Google's own ecological environment.

Reference Links:

[1]https://www.zdnet.com/article/google-tensor-everything-you-need-to-know-about-the-pixel-6-chip/
[2]https://techcrunch.com/2021/10/19/with-tensor-google-bets-on-its-own-chip-design-for-the-pixel-6/