Pixel 8 Pro reviews may be inadvertently convincing readers that the Tensor G3 is powering all new AI features ^↺

The Google Pixel 8 Pro has received many glowing reviews since its recent launch. However, as we have highlighted recent coverage, there are also matters of concern that have warranted closer scrutiny. We revealed that Google actively blocked reviewers from being able to freely and easily install benchmarks on their review units. We also drew attention to the Pixel 8 Pro review from YouTuber @Mrwhosetheboss, who revealed that many of the new AI and generative AI features Google has highlighted are not processed onboard by the Tensor G3, but off-loaded to the cloud instead.

Google’s narrative around the Tensor G3 processor is that it is an “AI-first” chip, and as a result, “traditional performance metrics” don’t apply. During its launch presentation — attended and viewed by many journalists from around the globe - Google highlighted a number of times how much faster and more capable the new Tensor G3 is than its predecessors at handling AI models. It also showcased a number of brand new AI features powered using AI algorithmic models. It was only when the company highlighted the new Video Boost feature later in its presentation that it made it clear that only some pre-processing was handled on board by the Tensor G3 before the video file was off-loaded to the cloud for the heavy lifting.

Not helping clear matters up, in an official Google blog, Monika Gupta, VP Product Management says of the Tensor G3 that “every major subsystem has been upgraded, paving the way for on-device generative AI.” As such, it is to see why a number of reviewers of the Pixel 8 Pro have drawn the conclusion that the latest AI features Google showcased are all running natively on the device. This misconception has extended to their published reviews, de-identified quotes from which we have included at the end of this article.

A common aspect of Pixel 8 Pro reviews that we have identified is that none we have read to date make a clear distinction between AI features performed on-device by the Tensor G3, and those AI features being off-loaded to the cloud for processing by Google's data centers. In one review, it was noted that it took a long time for Magic Editor finish processing a task. The reviewer was apparently unaware the video was being uploaded to the cloud first, where additional processing was taking place, before the finished product was made available. 

Only a handful of the reviews have made it clear - as per Google’s launch presentation - that the Video Boost function is being off-loaded to the cloud after some initial on-board processing. However, even in the reviews that pointed this out, none appeared to be aware that a number of the other new AI features Google showcased were also being off-loaded by the device to the cloud instead of being processed on-device by the Tensor G3. This is not surprising, as Google didn’t mention much of this during the launch presentation, nor does make much of this clear on the product page for the Pixel 8 Pro. As we now know, Google also admitted to Notebookcheck that it had blocked reviewers from being able to easily install benchmarks through the Play Store as they normally would.

Blocking reviewers from being able to easily install benchmarks is highly problematic as it stops them from being able to test the claims Google has made about its Tensor G3 chip against the objective data the benchmarks produce. Many of the same reviewers, even if they knew how to sideload the benchmarks manually, identified that Google’s Tensor G3 does not perform well compared to the competition. What these particular reviewers failed to note, however, is that AI performance of any chip is determined by its performance across the CPU, GPU in conjunction with its neural processor, which in Google’s case it calls the Tensor Processing Unit (TPU). 

A chip that is relatively weak across the CPU and GPU will also be relatively weak at on-device AI processing. This is why Google has been forced to off-load a number of these new AI features of the Pixel 8 Pro to the cloud, and also highlights why it is in fact very important the chips perform well in benchmark tests as it is also reflective of their overall ability to perform AI functions on-device. For example, GPUs on mobile SoCs are not just used by the chip for playing games, but they also happen to be exceptionally good at running AI models as they can run multiple machine learning threads simultaneously – also known as parallel processing. 

It is no coincidence, then, with the rise of generative AI over the past twelve months that Nvidia's stock has skyrocketed as its GPUs are often used in cloud servers for this very purpose. In fact, as per the Google Cloud website, Google’s cloud servers use Nvidia GPUs “on Google Cloud for machine learning, scientific computing, and generative AI.” This means when the Tensor G3 isn’t up to the task of performing AI functions locally, Nvidia’s GPUs in Google Cloud are doing all the heavy lifting for Pixel 8 Pro users. This also completely neutralizes the argument that raw performance in mobile chips doesn’t matter. It clearly does, and the rise of generative AI coupled with the inability of the Tensor G3 to be able to fully handle it on-device underscores this.

One has to feel for the “average user,” however, when it comes to all of this. If reviewers are unable to easily distinguish between features that are actually handled by the Pixel 8 Pro, how is the average user going to make sense of it? Especially when the Google marketing department is espousing how “super fast and efficient” the Tensor G3 is, and that it is "the power behind Google AI on Pixel". It is true they are buying an "AI-first" smartphone, but the reality is it is actually co-powered by Nvidia GPUs in Google Cloud.

Select quotes from various reviews of the Pixel 8 Pro from around the web:

Yet I have uttered the word "wow" multiple times while testing the Pixel 8 Pro over the past week, and the reason I'm so impressed is indeed that Pixel software — specifically, the Pixel 8 series' on-device generative AI capabilities. Every day for the past week, I've spent at least 30 minutes just sitting on my bed, playing around with them. 

XDA-Developers

 

The phone’s AI generates wallpaper from a series of themes and some simple prompts, and is capable of producing some stunning images. 

The Guardian

 

Alternatively, it’s interesting to note what’s happening in the background when Magic Editor is being used. That’s because Google AI is being leveraged to generate what it thinks should fill in the gaps. For example, I took a picture of my cat sitting on a small ottoman and decided to move him to the nearby sofa using Magic Editor…I can’t stop focusing on how the Pixel 8 Pro is one of the smartest phones I’ve ever come across. 

Tom's Guide

 

It's worth pointing out that you should be prepared to wait for some of these features as you use them. Generating photos in Magic Editor takes a good 20 to 30 seconds (horrid, I know), but Audio Magic Eraser can take even longer depending on the length of the video. Google also announced a feature called Video Boost (which is only on the Pixel 8 Pro) that lets you send a clip off to its servers to be processed. The cloud-based process improves stabilization, balances colors, reduces noise, and improves the brightness of low light footage. However, this processing could take several hours. 

WIRED

 

Features like Magic Editor are made possible by the new Google Tensor G3 chip, which powers standard tasks like app browsing, multitasking, and watching videos, and AI-enabled services like Google Assistant and Google Photos…These are the first phones that truly demonstrate how AI can help you throughout your day… 

ZDNET

 

And now on the Pixel 8, between the Tensor G3 and compelling new tools that chip is powering, Google is using AI to level up its playbook…Then we come to the Pixel 8 Pro, which is even more impressive. Google’s phones already took some of the best photos you could capture with a pocket-sized device. And with the power of AI and a full suite of upgraded sensors, the Pixel 8 Pro is basically unmatched when it comes to overall photography prowess. 

Engadget

For the record, and to the best of our knowledge, these are the new AI features processed on-device by the Pixel 8 Pro: Summarize for Recorder, Magic Eraser, Zoom Enhance, Best Take and Gboard Smart Replies.

These are the new AI features that are processed in part, or in full, off-device in Google Cloud: Google Assistant with Bard, Video Boost, Night Sight Video, Gboard proof-read, Magic Editor, AI Wallpaper, Audio Magic Eraser.

Purchase the Google Pixel 8 Pro from Amazon