Google's Tensor: The Data Company's Data Chip
Monday, November 15, 2021
Ever since the release of the iPhone in 2007, Apple has designed and fabricated its own chips for its own devices. At the time, "owning the supply chain" or the vertical, was the way of controlling the full stack of hardware down to software of the manufacturing and distribution process. Since then, we have seen the practice known as economies of scale, for Apple to make more revenue on each phone sold.
An unrealized benefit at the time is that the creation of a device makers own chips, also allows for unique customization and experimentation of SoC's to differentiate themselves from each other. Other examples include Samsung utilizing its own chips overseas, Microsoft's SQ series chips in the Surface Pro X, Windows ARM offering, and the newest entrance: Google's Tensor chip, which is the focus here. What's important to take away from these examples is the ease of which the manufacturer owns both the hardware and software stack, so in theory, components can become efficient and more intelligent with the entire device. Outside of mobile, we see Apple bringing the same concept to its laptops with the M1 series.
As the focus of this blog turns to all things data, the Tensor chip is the most interesting and dynamic from the aspect of AI and ML on the new Pixel 6 and Pixel 6 Pro devices. This is not a site for phone reviews, so I will not stray into its review, but rather what the Tensor's specifications and future is for Google.
Throughout the pandemic, Google was slow to release innovative instances of its Pixel line. Likely due to chip shortages and creating a SoC from scratch, Google entered the market last month with the Tensor, which is unlike anything else on the market, for better or worse. Since Google, itself, is not a semiconductor company, nor does it utilize fabs, they have chosen Samsung to produce the final product.
Functionality such as improved speech recognition, voice typing, live translate, and magic eraser to remove photobombers from pictures are all based on AI.
ZDNet
As the Tensor based devices offer these differentiations on product use, one of the often-forgotten benefits to a AI/ML blend is the security on the SoC. The Titan M2 component on the die, allows for hardware based security that will ideally stop any attacks aimed at the device itself; i.e. brut-force entrance by bypassing the fingerprint sensor.
Google's first Tensor device will learn from your habits and make suggestions based on usage to save battery, utilize Automatic Speech Processing, and the magic eraser to get rid of those unwanted background intruders on your photos. Out of all of the Pixel's features over the years, the place where the Tensor SoC really shines is its computational photography.
Other cool camera features thanks to Tensor include Magic Eraser, a feature that erases unwanted objects or people from photos. This feature uses Google's ML to do the task of what somebody would need to do on Photoshop, but in an instant.
Tom's Hardware
Rather than optics like a traditional camera, Tensor utilizes it's AI components to fill-in areas that are dull, or even missing. In theory, the machine learning component of the SoC, will allow features like Magic Eraser and Face Unblur to improve over time based on individual usage trends.
Given that this is the first generation Tensor SoC and Google is primarily a data company, this type of component is its core competency. Though Google is famous for deprecating or ending previous products like Google Wave, and Google+, research intensive projects such as SoC design and implementation is not something that can easily be explained away as a "failed software product". Hardware costs much more to develop in any technology company's R&D department.
The rumor mill is already circulating about the next generation Pixel, presumably the Pixel 7, with the next generation Tensor stack. It would make sense given all of the data and usage Google has collected from the first Tensor chips and their usage, and utilize that to improve the AI & ML on future devices.
Disclosure: I own the Pixel 6 and use it as my daily driver.