What is the level of Google’s self-developed chip Tensor? | Hard Philosophy

Compatible with the inclusive Android system, it creates an Android smart phone that blooms in abundance. But with regard to the title of the King of Machines, everyone has their own answers in their hearts.

CNN (Cable News Network) gave the Pixel 6 Pro the title of "Currently Best Android Phone", saying that it has an excellent camera, a smooth system and a unique appearance.

▲ Android 12.

In addition to the Android 12 system, Google’s first self-developed chip, Tensor, has created these features. Such a combination of hardware and software has made Pixel, the most competitive and attention-grabbing Pixel, not one of them.

Before the Pixel was unveiled, the pearls with Apple's self-developed A-series chips were in the front, and Tensor was also labeled "Boom Field". It seems that Google will also come up with a self-developed SoC that is worthy of history.

Only with the shipment of Pixel, Tensor's architecture and performance have also been completely deconstructed, but its performance and power consumption are not enough to be considered "mature". However, in the selection of some functional modules, it is quite the ingenuity of Google's "design".

Is it self-study or "magic reform"

The semiconductor industry in the 21st century is not like the controversy of a hundred schools of thought in the 1980s, it is more like a three-part world between a few big oligarchs.

The Google Silicon team has been investing in Tensor for three to four years , but for a processor SoC for smartphones, there is no deep technical accumulation. It is a bit of a dream to "debut is the peak".

▲ Google Tensor. Picture from: @Sundar Pichai

Apple next door has invested in the chip industry for nearly 30 years. Aquarius and PowerPC both lost ground. Finally, they released self-developed A4 chips in 2010, and finally made a big hit on A7, leaving several opponents behind. Body position.

The Google Silicon team did not have previous experience in the design and manufacturing of complex SoC chips. Previous works were more of the image processing chip PVC of Pixel 2, Pixel 3, Pixel 3 and subsequent Titan M security chips.

▲ Titan M security chip.

Tensor is the first "work" of the Google Silicon team.

Phil Carmack, the vice president of the Google Silicon team, said frankly in an interview with Ars Technica , "Although we are a team that is new to the SoC field, we know how to build a professional chip. We have Very reliable implementation path."

Therefore, some Tensors similar to Cheng Yaojin are self-developed, which has attracted considerable attention. After the Pixel 6 Pro was released, Anandtech conducted a thorough deconstruction and analysis of Tesnor.

To put it simply, the naming rules of Tensor and Samsung Exynos are relatively close, and a few years ago, it was reported that Samsung Semiconductor had begun to provide services for "semi-custom" chips.

ETNews even stated in its August report that Samsung will provide customized technology and functions according to customer needs, and it can even be involved in the chip design stage.

▲ Google Tensor silk screen is internally named S5P9845, while Exynos 2100 is S5E9840. Picture from: TechInsights

Although Samsung’s Exynos chips are not as good as Qualcomm and MediaTek in recent years, Samsung is no longer a simple chip manufacturer in terms of role, and the pattern has opened up.

In addition to background information, Tensor and Exynos use the same CPU and GPU architecture, and the large functional modules such as power management, IO, storage controller, etc. are all the same. However, there is a shadow of "customization" on the design of the large functional modules of the SoC. Tensor is very different from Exynos.

▲ Samsung Galaxy S21 Ultra and Google Pixel 6 Pro. Picture from: digitaltrends

To put it bluntly, Tensor is a customer of Samsung's "semi-custom" chip service, Google provides the design theme, and Samsung is responsible for the construction and production (here is a foreshadowing).

Strictly speaking, Tensor should be a "customized" chip, not created from 0 to 1, but from 1 to 2, or 1 to 3.

This also basically solved a mystery of Tensor's life experience, but this does not negate the investment of the Google Silicon team. After all, the chip is not a work of art.

What is the level of Google Tensor?

If you can summarize it in one sentence, Google Tensor's CPU is probably A12 (which is dragged down by the mid-core), GPU is stronger than Qualcomm Snapdragon 888+ (but the power consumption is extremely high), and AI performance is unprecedented (3 times 888+).

Tensor's CPU did not choose the mainstream 1+3+4 architecture, but chose the 2+2+4 architecture, two X1 large cores, two A76 large (medium) cores, and four A55 small cores.

▲ Exynos 2100 and Google Tensor of the same clan but different sources. Picture from: Anandtech

Specifically, Tensor's X1 core 2.8GHz frequency is lower than Exynos 2100 and Qualcomm Snapdragon 888 (888+), and compared to Exynos 2100, Google designed a 1MB L2 cache, which is equal to Snapdragon 888. More than Exynos 2100.

The two frequency-reduced X1s are unmatched in performance, and the lower frequency selection also allows Tensor to run high loads for a long time without frequency reduction.

When it came to the large (medium) core, Tensor did not choose the updated Cortex-A78 core, and used Cortex-A76 next to it. A76 is actually the core architecture of two years ago. It first appeared on the Qualcomm Snapdragon 855, but now it mostly appears on the Snapdragon 7 series (here is another foreshadowing).

The small core uses Cortex A55 with a frequency of 1.8GHz, which is standard for high-end chips. And Google also equipped with twice the L2 cache of Exynos 2100, which came to 128KB, again in line with Snapdragon 888.

Similar to the large core A76, the small core A55 also has a "mystery". Tensor binds the L3 cache to the A55 core frequency, which is different from the dedicated L3 cache frequency of the Exynos 2100, which will cause latency and power consumption problems (plus one more Foreshadowing).

On the GPU, Tensor is equipped with Mali-G78 MP20, which is almost the ceiling of the public version of G78. The L2 frequency is directly pulled to the frenzied 996MHz. Compared with the Exynos MP14, it has increased the number of cores by 42% and also increased the frequency. , That is to trade power consumption for extreme performance.

The actual performance of the Tensor designed in this way is actually a bit biased. The existence of 2 large cores makes its single-core performance sufficient, but the existence of A76 drags down the multi-core performance of Tensor as a whole.

▲ Google Tensor GeekBench 5 running points. Picture from: anandtech

However, the most fatal thing about Tensor is its high memory latency, even inferior to Exynos 2100. While the CPU is waiting for the memory, it is constantly glowing and heating. In a round of testing, Anandtech stated that Tensor took longer and had lower scores than Snapdragon 888, but it consumed more power (13.8% more).

Among the CPUs, there are only two X1 cores that are eye-catching, and the energy-efficient A76 and the A55 bound to the L3 eventually lead to the Tensor heating and slow operating efficiency.

GPU is like theoretical analysis. It has high energy and high consumption, and its peak power is directly piled up to 8~10W. However, Pixel 6 and 6 Pro have not used the VC heat sink or other heat dissipation methods of the current Qualcomm SoC Android phone. The heat dissipation level is "more like an iPhone. It’s not Android.” The heat buildup is severe and can only be reduced frequency. Before completing a round of testing, Pixel started to reduce frequency.

As for the GPU rushing to the power level of 8~10W, I have only seen it in some Snapdragon 888 e-sports game phones. You must know that they have a PC-level active cooling system and an external power supply.

To say that there are some regrets in the design of CPU and GPU, then in the configuration of ISP (image processor) and TPU (machine learning engine), Google has thoroughly exerted its advantages.

ISP integrates Exynos and Google’s custom image chips. The Exynos part is responsible for image collection and pre-processing, while the Google custom part is responsible for "calculation", which is reflected in the Pixel 6 Pro product, which is the video HDR Net, dynamic Blur, character keying, etc. functions.

And Tensor's built-in TPU is the reason for Tensor's name. It uses the latest machine learning processing architecture and is optimized for the Android 12 system. However, like the machine learning and AI modules of other chips, the features that can be directly expressed are still relatively " Vague".

The most intuitive features on Pixel 6, 6 Pro are the system, smooth animation, and fast voice and image recognition. As for the performance of this TPU on Tensor, in the running scores of some models, it also surpasses the popular mainstream SoCs (including Qualcomm Snapdragon 888 and Exynos 2100).

In addition, Google has not officially announced the SDK of this TPU to developers, so the powerful TPU built in Tensor is still exclusive to Google, temporarily lagging behind the development ecology of Apple's A-series chips.

Why is Google Tensor designed like this?

When deconstructing Tensor, I left a few foreshadowings. One is Samsung's production, the other is the choice of the CPU core A76, and the third is the design of the A55 core.

In fact, there are still many questions about the design choices of Tensor, such as the choice of 2+2+4 architecture, the original intention of Tensor and so on.

▲ Monika Gupta, Senior Director of Google Silicon Team. Picture from: businessworld.in

"For Google, we want to apply AI to all aspects of our lives," said Monika Gupta, senior director of the Google Silicon team. AI is designed according to our preferences.”

"We don't want to produce smartphones in the traditional sense." Google wants devices with a greatly increased processing power of AI and machine learning, but the processors on the market do not meet Google's needs. Google chooses self-developed chips for stronger AI. The original intention .

This is also the reason why we see a sufficiently powerful TPU on Tensor. The most commonly used name in the Google AI department is "Tensor", and the name Tensor is quite obvious. In addition, the word Tensor itself is very Google, and it has an engineer culture.

"The machine learning code of Pixel 6 can still run on the old Pixel, but the efficiency is a bit worse." The unique AI performance in Tensor is not "empty talk", Monica added "Google has put the latest and greatest R&D department Strong results are present in Pixel 6, 6 Pro."

"Everything is consistent with the purpose you want to achieve." This purpose is efficiency. Google believes that the efficiency of two super large cores running at medium load is much higher than that of a large core, and the energy efficiency ratio is also higher.

Phil Carmack, the vice president of the Google Silicon team, also gave an example, "Turn on the camera, in addition to everything you see, the CPU, GPU, ISP, and TPU inside the SoC are constantly running, calculating, and complicated. The scene of ”will involve a lot of machine calculations.” At this time, two X1 processors are handed over, and it is easier to do.

"If you need more flexible response speed, high efficiency, and high performance to achieve your goals, the two X1 equipment is better than the current super-large core equipment."

As for the big core's choice of the 5nm A76 instead of the updated A78, it did not appear in this interview. Therefore, we can only guess Google's intentions. There are probably such several situations.

First, Google believes that the performance of the A76 under the 5nm process is 20% higher than that of the 7nm A76 a few years ago, which is enough to cope with a light load, and the higher it is, it will be handed over to the double X1.

The second is that Google has a wealth of experience in tuning the A76 core, which is more conducive to directly reusing the machine learning previously on Pixel without re-adapting the call.

Thirdly, when Google and Samsung were cooperating to design SoC, because the A78 was too new, Samsung was unable to provide the A78 core for the time being, and Google retreated to the A76.

Regardless of the considerations, the first generation of Tensor still gave Google a bit of a bite in the selection of CPU and GPU. With the current Android 12 scheduling, the 2+2+4 architecture still has a certain gap with Google's expectations.

Why is Google making cores?

Google CEO Sundar Pichai called the Tensor chip "the biggest innovation in the Pixel field so far."

▲ Pixel has undergone earth-shaking changes. Picture from: CNET

For Pixel, even if Tensor is aside, Pixel 6, 6 Pro are still products that Pixel opens a new chapter, returning to mainstream hardware, brand new Android 12 and enough Google design.

The "self-developed" Tensor has not surpassed the current generation of processors, but under the design of Google, it seeks to solve the problem in its own way and presents an unconventional mobile phone.

Even though the Pixel 6 and 6 Pro with Android 12 have not yet performed as they should, the AI ​​and machine learning functions emphasized by Google are temporarily not as powerful as imagined. However, it is only a matter of time before Google, a major AI and machine learning company, thoroughly understands Tensor.

As the Google Silicon team Carmack said, under a unified processor architecture, there are too many traditional products on the market. Google's Pixel series wants to get a piece of cake. It is a good choice to choose different strategies.

Not only Google, mobile phone manufacturers have become a trend in self-made core, such as vivo V1, Xiaomi's surging, in essence, closer to Tensor, it is a unique and distinctive product.

And I want to have enough customization rights for the product, instead of constantly adjusting the product form with the mainstream processor architecture in the market, just like this year's Android flagship all has the internal heat dissipation context.

In this sense, Tensor, whose absolute performance is not in the first echelon, is actually a success for Google. It has not been engulfed by the market tide. Tensor's AI and machine learning design has made the Pixel series the strongest in Google. Imprint.

#Welcome to follow Aifaner's official WeChat account: Aifaner (WeChat ID: ifanr), more exciting content will be provided to you as soon as possible.

Ai Faner | Original link · View comments · Sina Weibo