The Google Tensor chip is undoubtedly the Pixel 6’s greatest curiosity. Anandtech has peeled it all over to better understand Google’s options.
we tested Pixel 6 And 6 ProAnd we took a closer look at the Tensor chip. As a reminder, the Pixel 6 integrates for the first time a home SoC designed by Google called Google Tensor Simply. It is said that Google plans to develop one every yearThey can even create more powerful chipsets for Chromebooks for example.
Anandtech Media Specialist Take a closer look at this Google Tensor slide, They peeled it all over.
A slide designed by Google and Samsung
As expected, the Google Tensor SoC is derived from Samsung Exynos 2100 that we find in Galaxy S21. So Google worked alongside Samsung to design their chip, and there are many points in common between the two.
|Google Tensor||Qualcomm Snapdragon 888|
|CPU||2x Cortex-X1 @ 2.8 GHz
2x Cortex-A76 @ 2.25GHz
4x Cortex-A55 @ 1.8GHz
|1x Cortex-X1 @ 2.8GHz
3x Cortex-A78 at 2.4GHz
4x Cortex-A55 @ 1.8GHz
|GPU||Mali G78 MP20||Mali G78 MP14|
@ 1066 MHz
@ 1352 MHz
|Internet service provider||Hybrid Exynos + Google ISP||Exynos ISP full blocks
|modem||Exynos 5123 . modem||Exynos 5123 . modem|
As you can see in the table above, the ISP (image signal processor) or even the 5G modem are identical on both chips. This does not mean that the chips are the same, as you can see in the table, the CPU and GPU are different. Google used two powerful cores, the popular ARM Cortex-X1 instead of the one on the Exynos. In fact, technically, the Tensor CPU is closer to the Qualcomm Snapdragon 888.
Google Tensor CPUs and GPUs Don’t Shine
In Anandtech’s tests, all three chips are very similar. The graph below is quite interesting as it compares the power to the power consumption of each CPU core.
And so we can see the Cortex-X1, but also the Cortex-A76. In overall results, the Apple chips, A15 and A14, are more efficient. As for the performance of the Tensor / GS101, it is slightly lower than that of the Exynos 2100 and lags behind the Snapdragon 888 by a noticeable margin (12.2%), while it consumes 13.8% of power. This difference is related to memory management by the Tensor chip, as well as a higher latency. This can also be explained by the cooling system, which is probably less efficient as we will see below.
Although Google Tensor has twice as many X1 cores as Qualcomm and other Samsung SoCs, the Google chip does less work on GeekBench (graph below). This is because Cortex-A76 cores, the mid-range cores, are worse than competitors. This can also be seen in the above graph.
The difference with Samsung and Qualcomm will be more on the GPU side. We find an ARM Mali-G78 GPU with 20 cores with 42% more cores than the Exynos 2100.
In practice, the results obtained by Anandtech are very average. On paper, the massive G78MP20 GPU looked like a tyrant, but in practice it falls short of theoretical performance. The graphics performance of the Google Tensor chip is no better than that of the Exynos 2100, moreover, it is also hampered by the design of phones.
In fact, the US media reports that the Pixel 6’s internal design is closer to the iPhone than to other smartphones, since the SoC is not glued to the screen. It is better for repair, it is worse for heat dissipation. So Anandtech has measured points of extreme heat, especially on the left side of the phone near the SoC.
Unique neural engine
The completely different component is the neural motor. This term covers several components, including TPU (Tensor Processing Unit) and NPU (Neural Network Processing Unit). As an anecdote, it was TPU that gave the Google chip its trade name.
Here, Anandtech highlights Google’s great work on TPU design. This, in their opinion, is a very impressive competitive advantage over other platforms, as it represents a huge leap in performance.
This explains the results we’ve seen about the Pixel 6’s photo capabilities, for example, compared to competing smartphones. The graph above shows just one aspect of this neural engine, the Google Tensor chip performs better than Apple in many areas, including language analysis for example or image recognition.
Google designed a homemade chip to differentiate itself
We better understand why Google designed a homemade chip for its smartphones. They felt that they could create value by choosing a personalized approach to achieve better performance and efficiency in AI-related tasks.
This work by AnandTech also shows the collaboration between Samsung and Google, while the US company did not mention this in its communications. Some of the hallmarks of Exynos chips are found in Tensor, especially energy efficiency. This means that the Tensor chip, just like the Exynos, lags behind Qualcomm’s Snapdragon solutions when it comes to battery life.
The same goes for choosing a larger GPU. In theory, one can expect better performance. In practice, the pixel design is detrimental to heat dissipation. This first coffee is not a disappointment, however there are design weaknesses and some options can be criticized.
In the end, the neural processor is the only area where Google Tensor really shines. Natural language processing seems to be the strong point of Google and Tensor, with undoubtedly a big lead over the competition. It will be interesting to see how Google develops its future chip, the Tensor 2, to differentiate it from Samsung’s Exynos.
“Incurable web evangelist. Hipster-friendly gamer. Award-winning entrepreneur. Falls down a lot.”