Google Unveils The New Tensor SoC, The Heart Of The New Pixel 6 Line Of Phones

Google Machine Learning Engine GOOGLE

It started as rumors and was eventually confirmed by Google that the company would be replacing Qualcomm’s chipsets in favor of its custom silicon. Some suspected that Google’s design would be completely new, while others suspected it would be some variation of Samsung’s Exynos processor. While it appears that Google is using a Samsung 5G modem in the new Pixel 6, the Tensor processor appears to be a newly configured Arm processor using Arm cores but still borrowed from Exynos. Unlike the situation with the the Microsoft Xbox processor where Microsoft gave AMD credit, Google took all the credit without ever mentioning Samsung as if Samsung had nothing to do with it. 

Google claims that the company’s desire to create the Tensor SoC was because it was innovating faster than the industry could provide chips. This seems like an odd claim because last year’s Pixel 5 shipped with a Snapdragon 765G, a considerably less capable chip than Qualcomm’s Snapdragon 865 flagship processor last year. This is relevant because, in many of Google’s performance claims for the Tensor SoC, Google compares against the Pixel 5 and Pixel 4. 

During the Pixel 6 launch, Google talked about how vital ML performance is to the company and that not any single benchmark of any specific component matters as much. But this seems to contrast Google’s decision last year to ship the Pixel 5 with the Snapdragon 765G, which only had 5 TOPs of AI performance compared to the Snapdragon 865, which had over 15 TOPs or triple the AI performance. The Snapdragon 855 before that in the Pixel 4, which Google compares to claimed ML performance of 7 TOPS, explained why some users (Anshel included) noticed that the Pixel 5 felt sluggish in some AI tasks like noise reduction and other image processing. 

The new Tensor SoC features two Arm Cortex X1 cores, clock speed unknown but likely lower than contemporaries, paired with two Arm Cortex A76 cores, and a quad-core complex of Arm A55 little cores. To my knowledge, this configuration has not yet been done by any of Arm’s partners, as Samsung and Qualcomm have both gone with a single X1 Arm core and then used three A78 cores and four A55 cores to build their SoCs. Additionally, Google’s Tensor SoC uses Arm’s Mali G78 GPU cores, which are the same as in the Exynos 2100; however, it appears that Google uses a 20 Core GPU design compared to Samsung’s 14. Google manufactures The Tensor SoC on Samsung’s 5 “nm” process node, the same as Samsung’s leading Exynos SoCs today. That said, this design does appear to be more custom than a simply re-badged Exynos 2100 chip, but at the same time, outside of the TPU and some of the security elements, this does appear to be a primarily off-the-shelf Arm design designed by Samsung with similar Exynos building blocks, just more or different frequencies. Google has chosen to go with 4MB of CPU L3 Cache on this design with 8MB of system-level cache. The CPU level cache is identical to the Exynos 2100, while the system cache is slightly increased from 6MB. Extra system level cache could potentially help the Tensor’s performance in single-threaded tasks and memory bandwidth.

We’ve seen Google use all of the comparisons in its documents about the Tensor SoC compared to the Snapdragon 765G and Snapdragon 855, both of which do not represent the best of what’s available from the competition. It will be interesting to see how well other SoCs fair against the Tensor SoC in ML benchmarks and whether Google’s efforts will gain an edge in AI against the competition. With two X-1s likely operating at lower frequencies I could see it lagging in single threaded performance which drives snappiness of apps. Others are saying it will be at the same frequency , do the same on ST and do better with MT. With that said, nobody can definitively talk about performance quite yet, so we will have to wait and see. That also goes for the modem inside the Pixel 6, which is still rumored to be a Samsung 5G modem like what Samsung ships globally.

The still unconfirmed modem inside the Pixel 6 appears to be the first time Samsung has ever commercially shipped mmWave in any region. As far as the Pixel 6 goes, only the US and Japan are getting mmWave, so the rest of the world doesn’t even get a chance to experience this capability. That said, the US, mostly Verizon, is the primary reason for mmWave support today, but I believe that long-term mmWave will be valuable in most countries with densely populated areas. While it remains to be seen how this 5G modem performs in both throughput and power consumption, it will be difficult to tell until users get their hands on the modem and we get some teardowns to tell us about the modem, mmWave modules used, and which generation of Samsung’s modem Google is using. 

After launch, Google caused some confusion about the capabilities of different models, including the Pixel 6 and Pixel 6 Pro, with a spec sheet that contradicted the official specs of the Pixel 6 lineup. Google is only shipping mmWave in the US for the Pixel 6 for AT&T and Verizon, while T-Mobile only gets 5G Sub-6. Additionally, all US versions of the Pixel 6 Pro get mmWave and Sub-6 support, and the Japanese variant also gets mmWave and Sub-6. Outside of these regions and carriers, the rest of the world only gets Sub-6 5G capabilities.  What’s interesting, though, is that the Verizon Pixel 6 with mmWave sells for $100 more than the T-Mobile variant. The AT&T version of the Pixel 6 with mmWave ships for an even more expensive $740, making the $899 Pixel Pro look a lot more attractive with the larger display, higher resolution, mmWave 5G, and better camera configuration. Confused yet about the new Pixel and 5G support? I am, and I think any buyer that cares about 5G will be confused, too.

Google has put a lot of effort into building the Tensor SoC, and it remains to be seen where else this SoC may land, as I suspect that Chromebooks won’t be far behind the Pixel lineup, especially considering the specs of this chip. Google’s current global market share with the Pixel line of phones is just around 1-0.3%, depending on what data you’re looking for, which isn’t much compared to many of Google’s competitors. It’s also not likely to dent Qualcomm’s financial anytime soon and would take a massive share shift to ever dent it. Building and maintaining a custom SoC is not a trivial pursuit; it requires many engineers and fab partners and big upfront costs, including tape-outs that on leading nodes can run as high a $50 million per chip, not to mention licensing fees. Unlike Apple, Google also has to pay Samsung’s custom SoC group as well. In general, it costs hundreds of millions of dollars to design and develop an SoC and add more for your custom IP and license fees to companies like Arm. Until Google finds a way to sell significantly more Pixel phones or expand the reach of the Tensor SoC line, it may be a sunk cost that the company may never recover. Companies like Apple and Samsung can build their custom chips because they ship enough phones with their custom chips to justify the cost and amortize those development costs across hundreds of millions of devices every year. I believe Apple has 75% as many engineers working on its SoCs than AMD does today. Is Google willing to invest what Apple and AMD are to lead with its own SoCs? We’ll see. Also, let’s not forget that Apple and AWS’s custom chip development took 10 years to master and Google is beginning its journey today. Don’t confuse Google’s datacenter TPU either with a mobile SoC as the designs and skills are valleys apart. I do wish Google the best of luck as I have always said, more competition is better.

Note: Moor Insights & Strategy principal analyst Anshel Sag co-wrote this article.