Among the string of announcements that were made at Google I/O 2022, most of its new devices, including the mid-range Pixel 6a, the Pixel 7 and Pixel 7 Pro and the Google Pixel tablet, will all be run using one chip – the tensor. Last year, the launch of Google Pixel 6 and Google Pixel 6 Pro phones marked the first time the tensor chip was used in Google’s devices. Even alongside the Pixel 6 and Pixel 6 Pro phones, it was Google’s tensor chip that became the most important release last year. Not only did the system-on-a-chip or SoC add features like live translation and faster computational abilities, it also hinted at the tech giant’s ambitions to expand strategically as a hardware manufacturer. At the Pixel launch event last year, the company called the tensor the “biggest mobile hardware innovation in the history of the company.” If there were any doubts about whether Google had bigger plans for its tensor chip, this year’s Google I/O conference showed that the answer was a resounding yes. 

Why tensor?

In pursuit of ever-present technology, ambient computing and pushing AI-backed features, the company realised that it couldn’t simply build a single co-processor like a Pixel Visual or a Neural Core to work with its AI models. Instead, the new Pixel model would need a whole new chip that could optimise tasks. Using the Exynos SoC as a template, Google collaborated with Samsung to build the tensor. 

Google said that the tensor’s CPU performance was 80 per cent faster than the Snapdragon 765G from Pixel 5, while the GPU performance was faster by 370 per cent due to the 20-core ARM Mali-G78 MP20. The tensor had a slightly different configuration. While it used the same cores built on Samsung’s 5nm LPE manufacturing process, tensor employed four A55 cores for background tasks and two big A76 cores instead of one. 

Source: Google

Google’s Tensor Processing Unit (TPU), traditionally called the Neural Processing Unit (NPU), is where most of its AI and ML operations take place. This usually is the part responsible for a bunch of Pixel 6 features, like live language translations, automatic speech recognition used in live captions, and photography that uses computation, like in motion mode and long exposure. While people from the modding community have attempted to activate Pixel 6 features on other devices, it either takes much longer or warms up the phone too quickly. This means that Google was justified in manufacturing the tensor chip because otherwise, the features would slow down other phones without optimised hardware. This is why the real benefit of the tensor chip is actually derived from Google’s custom cores. 

Image processing and Assistant voice typing features with the tensor

The tensor chip also includes an improved Image Signal Processor (ISP) chain. Tensor’s ISP chain works on turning raw data into pixels, interfacing with cameras and for noise and colour correction. With Pixel 6 and Pixel 6 Pro, Google introduced two new features – Real Tone, which represented skin tones as accurately as possible, and HDRNet, for real-time processing of videos. Meanwhile, face detection is faster and more efficient with Pixel 6 than with Pixel 5.

Since the launch of the Pixel 6, the Titan M2 was used as its dedicated security chip, making it Google’s most secure phone released until then. Aside from protection against user data, the Titan M2 also shields the phone against electromagnetic analysis, voltage glitching and laser-fault injection. 

Peer pressure

Google could also have been possibly forced to manufacture the tensor chips after a competitive push from Apple, which switched to custom silicone in its phones. Apple was able to cut costs of components and boost the performance of its iPhones. Once this move showed success, Apple implemented the same with its Macbooks. With the current share that the tensor commands, it is difficult to say if Google will be able to save input costs. 

Breaking and entering the SoC market

Considering that the tensor chip has only been around for a brief amount of time in the market, it is surprising that it has already left a dent in the high-end chip market. A report published by Counterpoint Research showed that while the companies on the top of the heap were predictably Qualcomm and Samsung, Google’s tensor also featured on the list. Since the data was only taking 2021 into consideration, tensor technically only reflected its numbers for the last two months of last year. Despite this, sales of Pixel 6 and Pixel 6 Pro drove tensor to capture between 1 and 2 per cent of the high-end chip market. 

While it is not the simplest task to cement oneself as a leading market player among chipmakers, it looks like Google’s tensor has set the foundation correctly. 

The post This step by Google just made its tensor chip so ‘important’? appeared first on Analytics India Magazine.