The semiconductor industry is a crazy place.

There’s an ebb and flow in computing. A new chip is rarely a perfect one-to-one upgrade over an older chip. Different features, different initiatives, changing consumer needs, numerous factors influence a chip design.

Google’s Tensor is an interesting beast.

Tensor is an attempt at focusing phone hardware to better run Google software. It’s an attempt at creating more hardware and software synergy for Google services. It’s a surprisingly successful solution for a first generation System on Chip (SOC), but it is first gen and we should expect improvements from succeeding generations.
What does Tensor 2 need to be successful? Not a lot…

What Tensor Got Right

It’s my belief that Tensor 1 is a year out of its expected  launch cycle. I have no inside information to confirm this, but looking at the design of the CPU and GPU, I believe Tensor was originally conceived for a late 2020 launch. If anyone’s interested in my ramblings on SOC design, I can expand on that idea in another editorial.

The TLDR for my crazy CPU conspiracy theory: Tensor 1 competes surprisingly well against chips made in 2022, but it was likely built to compete against chips released in 2021.

Google’s claims of AI and improved image processing are well represented.

Tensor (like the Neural Core in the Pixel 4 before it) demonstrates some incredible performance when chewing up image data. Running the same basic video editing test I’ve clocked for years now, the Pixel 6A set a new record in trimming down a single video file, even outpacing the monster powerful Xiaomi 12S Ultra.

Where the Tensor loses fights, especially in GPU performance, it loses by a margin we would expect from a one year cycle looking at Qualcomm processors.

Google set out to build a high performance “engine” for their products, and considering this as a first attempt (with a somewhat unconventional design) it’s a fantastic achievement.

What Tensor Got Wrong

Well, Tensor is an unconventional design.

There’s been a trend in phone SOC design towards ONE BIG POWER CPU core, then some medium sized cores, and lastly some low power cores. That’s the current configuration for MediaTek, Samsung, and Qualcomm flagship SOCs.

Tensor 1 uses TWO BIG POWER cores, two smaller medium cores, and then four of those low power cores. Samsung is responsible for the SOC manufacturing, but this is not how Samsung makes their own chips for the Exynos line of SOCs.

It’s in this partnership between Google and Samsung that some eyebrows might be raised. Google isn’t leaning on Qualcomm for components, and Google isn’t using TSMC for the chip fabrication.  Qualcomm “switched horses mid-race” this year with the 8 Gen 1 SOC. Starting the year with Samsung, and ending the year with TSMC.

The Qualcomm chips fabbed by Samsung at the beginning of 2022 were incredibly powerful, but had a rather poor performance per watt. The Qualcomm chips fabbed from TSMC deliver similar (sometimes slightly better) performance, but are noticeably better behaved for power draw.

Samsung has struggled with their chip manufacturing. Yield and efficiency are still concerns, which can keep component prices high. Rumors are swirling that Samsung might take a year off the Exynos line in 2023, and only ship Qualcomm powered Galaxy phones, while they sort some of their manufacturing issues.

If that rumor proves true, Tensor would be one of the main “crown jewel” products delivered from Samsung’s foundry in 2023.

The other main issue for Tensor has been radio performance.

It’s really difficult beating Qualcomm at this telecommunications game. Qualcomm modems are fantastic.

[Brief tangent: Even Apple is back to using Qualcomm modems on the iPhone. Apple tried to circumvent Qualcomm by using Intel modems for several years, and they were noticeably poorer performers than the Qualcomm modems of the same era. Just a funny thing to remember, that when people say modem performance is some kind of “deal breaker” for using a Pixel, I wonder how many of those same reviewers ever said an Intel radio in an iPhone was an outright DO-NOT-BUY deal breaker.]

Having a poorer performing radio in a phone is a double whammy. The phone works harder to find signal. Struggling with reception in weak network conditions means slow data speeds or dropping calls. At the same time, since it’s working harder to find signal, it’s also running warmer and eating more battery.

Again, considering the “first gen” status of the Tensor, the radio is a respectably strong performer for outright data speeds. It just doesn’t live in a vacuum. Older Google mid-ranger phones can often hold on to signal better when using a Qualcomm radio.

If you live in good network coverage, then the radio on the Pixel 6 is going to be perfectly fine for most needs. It’s when you travel outside good coverage, those reception issues manifest.

Radio and display are two of the biggest areas where regular daily use impacts battery life.

Moving Forward

Honestly, Tensor 2 doesn’t need to change a lot.

The main rumors point to a refined manufacturing process, a newer radio, and a more powerful GPU. That’s exactly what I want to see.

Google has never claimed any dominance in traditional CPU benchmarking scores, and the main focus of a Google chip is to run Google software.

Since the Snapdragon 855 (released in 2019), we’ve been in an arms race for more compute power, but several years on, efficiency has taken a backseat in many new products.

This situation is not unique to AndroidLand, where even Apple has had to shift focus on efficiency over outright performance improvements. The iPhone 12 thermal throttled worse than the iPhone 11 in game play tests. The iPhone 13 delivered a bigger battery to balance the 5G radio and powerful A15 chip. Now the A16 chip in the iPhone 14 Pro is only a modest upgrade over the A15.

People still tout Apple’s unlabeled bar graph marketing as proof of some significant superiority, but the real world use of our phones shows a MUCH smaller advantage. Apple’s most expensive phones enjoy a small lead against the current generation of more affordable Android devices.

This is a unique opportunity window for Tensor 2.

Qualcomm has announced a focus on efficiency for next year’s 8 Gen 2. The Apple A16 is a modest upgrade, likely focusing more on AI hardware. Google has already demonstrated some positive results in Camera and AI co-processors.

Tensor 2 getting a die shrink from 5nm to 4nm should hopefully deliver some power savings. An improved radio, even if it can’t quite match the current Qualcomm offering, would also improve battery life and connection consistency.

Tensor TOO Clever?

Lastly, as a plea to the folks on the Pixel hardware team, maybe it’s OK if you make Tensor 2 a LITTLE dumber?

One of my biggest gripes in general with current phone design, is trying to make the automated modes on hardware too clever. The phone will manage charging and battery draw. The phone will manage CPU performance based on previous user interactions. The cameras are all on full auto “just push the shutter”.

What we see from Sony, Motorola, and some gaming phones however, there is room to include some power user management features. Some of my issues with Tensor 1 would immediately be solved by giving more control over the performance of the phone.

Samsung and OnePlus were both caught rigging the performance of their phones for better benchmark scores, but the PRACTICAL result of that performance manipulation was better for thermals and battery life. The problem wasn’t a “Lower Power Social Media App Mode”. The problem was a performance limiter that the user couldn’t control.

Most of my daily interactions with a Pixel require only a tiny sliver of the compute power I know Tensor is capable of, but there’s no way to know what the phone is doing when so much of the operation is “IT SHOULD JUST WORK FULL AUTO”.

Giving me a battery mode that changes the CPU behavior would be a welcome addition. I enjoy the performance improvements on my Pixel 6A, but I miss the battery life of my Pixel 5A.

For folks on the “6” generation of Pixel phones, some are managing power by disabling features like Device Health Services and Digital Wellbeing.  It minimizes some of the background operation and battery draw, but those are also some features people might appreciate having.

Wrap Up

I’m really curious to see what Google delivers on the Pixel 7.

I’ve had a great time using the Pixel 6 Pro. The launch was a bit rocky, but what the phone has evolved into has been exciting to watch. There’s very little I would change, given the reality of electronics pricing.

I genuinely don’t think “MOAR CPU PERFORMANCE” matters much in this age of mobile app development. We’ve been overbuying CPU power since 2018, and most developers want their apps to run well on LOWER power phones. They’ll make more money if people on entry level phones can use their software well too.

A Pixel 7 with a better fingerprint sensor, a nicer ultra-wide camera, and a more power efficient Tensor 2? That phone would CRUSH.