When teasing the Pixel 6 at the beginning of August, Google has framed the preview very much around its first bespoke System-on-a-Chip (SoC). At the Pixel Fall event today, Google extensively described Tensor and called it “the greatest mobile hardware innovation in the company’s history.”

Why Tensor

Google’s stated goal when developing Tensor is to promote what is feasible on smartphones. The company wants to bring “AI breakthroughs straight to the pixel” and advance its vision of a technology that is always available, ie ambient computing

The former emerged from Google’s hardware department, which is convinced that pixels with AI-supported intelligent functions can be differentiated from competitors, while Google regards phones as the “central control unit of an environmental system”. At the Pixel Launch Event, Google spoke again in particular about ambient computing. The last time that occurred in any significant way was in 2019 Pixel 4 launch.

In an interview with The edge, Rick Osterloh said the work began in 2017 after realizing that Google couldn’t take a piecemeal approach – like building a single co-processor, e.g. B. Pixel Visual / Neural Core – to improve AI models. Rather, an entire chip is required that is optimized for the desired tasks.

The Tensor Chip is specifically designed to bring the latest Google advances in AI right to a mobile device. This is an area we’ve held back for years, but now we can start a new chapter in AI-powered smartphone innovation.

The Tensor CPU + GPU

At the Pixel Launch Event, Google responded to Tensor and explicitly advertised the inclusion of two powerful ARM Cortex-X1 cores with 2.8 GHz. There are also two “medium” 2.25 GHz A76 CPU cores The Google Silicon interview from Ars Technica They point out that they are based on a 5nm process rather than the 7nm original found in flagship phone chips last year. Four highly efficient / small A55 cores round off the CPU.

The Dual X1 approach enables Google to deliver more power on medium intensity workloads. In a normal CPU, the middle cores would handle such tasks as Google Lens visual analysis, but would be “maximum”. According to Google, using two X1 cores would be more efficient in this scenario, and that’s what Tensor is optimized for. In fact, it’s 80% faster than the Pixel 5’s Snapdragon 765G.

“You can use the two X1s at a lower frequency to make them extremely efficient, but they still have a pretty heavy workload. A workload that you would normally have fully exhausted with dual A76s now barely accelerates with dual X1s. “

Phil Carmack, VP and GM of Google Silicon

There’s also a 20-core GPU that Google says will “provide a premium gaming experience for the most popular Android games.” It’s 370% faster than the Pixel 5 that uses the Adreno 620 GPU.

Tensor safety ft. Titan M2

The security kernel of Tensor is a CPU-based subsystem that is isolated from the application processor and is dedicated to the execution of sensitive tasks and controls. It works with the dedicated Titan M2 security chip, which is not part of Tensor, but is considered by Google to be resistant to advanced attacks like. is touted electromagnetic Analysis, Voltage disturbances, and Laser defect injection.

the original The Titan M chip, in conjunction with software, prevents your phone from being rolled back to an older version of Android that may have security flaws. It also prevents the bootloader from unlocking and verifies your lock screen passcode.

TPU, ISP and Context Hub

Of course there is also the “Tensor Processing Unit”. This ML engine is said to be “tailor-made by Google Research for Google Research” and to build on where “ML models are on their way, not where they are today”.

The image signal processor (ISP) has an accelerator that runs the HDRNet algorithm, a big reason the Pixel 6 and Pixel 6 Pro can Live HDR + video at 4K 60FPS, more efficient.

The Context Hub brings “machine learning into the domain with extremely low power consumption”. It enables the Always-On-Display (AOD), Now Playing and other “ambient experiences” to run “all the time without draining the battery”.

Now all together or heterogeneous computing

All of these components together form a tensor, with Google prioritizing “overall performance and efficiency”. This includes in particular the outstanding performance in heterogeneous computing tasks in which different parts of the SoC have to work together. For example, Lens uses the CPU, GPU, ISP, and TPU to run efficiently.

As software applications become more complex on cell phones, they run on multiple parts of the chip. This is heterogeneous computing. To get good performance for these complex applications, we made system-level decisions for the SoC. We made sure that different subsystems within Tensor work really well together instead of optimizing individual elements for top speeds.

What tensor can do

In addition to Live HDR +, which makes colors more accurate and more vivid, Tensor enables other computer-aided photography and video functions such as the movement mode in Google Camera at 4K60. Action Pan blurs the background while the long exposure acts on the subject (see below).

Meanwhile, the face recognition on the Pixel 6 is more accurate and works faster – thanks to the integrated subsystems, while it consumes only half as much power compared to a Pixel 5.

Assistant on Tensor uses “the most advanced speech recognition model ever published by Google” with again half the performance. The high-quality ASR (automatic speech recognition) model is used for transcribing voice commands and in long-running applications such as recorder and live caption, “without quickly draining the battery”.

Meanwhile there are Assistant voice input for editing the content just transcribed completely freehand, and Live Translate with an 18% improvement in the translation quality of the pixel, “a level of improvement that normally requires several years of research”:

Google Tensor also enables Live Translate to manipulate media such as videos using language and translation models on the device. Compared to previous models on Pixel 4 phones, the new neural machine translation (NMT) model on the device consumes less than half the power when running on Google Tensor.

Tensor’s future

Google doesn’t give Tensor a generation identifier when it launches, but the company will likely append a number to the next version. (For example, the Titan M is being replaced by the Titan M2.)

There’s no doubt that Google is making more chips for phones (and others) Form factors are rumored). This is what SVP Rick Osterloh said at the event:

Tensor also provides us with a hardware foundation that we will build on for the years to come so that you can get the personal, helpful experience you have come to expect from a Google phone.

FTC: We Use Income Earning Auto Affiliate Links. More.

You can find more news at 9to5Google on YouTube:

Source link

Leave a Reply