Why has an AIoT chip become a battleground?

In recent years, the term AIoT has gradually become popular in people's vision. The so-called AIoT, namely AI + IoT, means combining artificial intelligence technology with Internet of Things technology to form a new integrated discipline.

With the development of artificial intelligence technology, traditional IoT devices will tend to be intelligent, thus forming the AIoT artificial intelligence Internet of Things, enabling the "Internet of Everything" to evolve into the "Intelligent Internet of Everything".

Of course, the development of technology is inseparable from market demand. As people face more and more demands for intelligent applications, AIoT faces more complex scenarios, and people need to provide it with more efficient computing to achieve faster responses.

Against this background, AIoT chips came into being. When facing AIoT scenario tasks, they have more advantages in computing power and power consumption than traditional general-purpose chips.

What is AIoT chip?

What is an AIoT chip? As we all know, IoT devices are different from mobile communication devices such as mobile phones. IoT devices are diverse, facing different application scenarios, and have different requirements for AI computing power.

At present, a universal chip architecture that can exist across device forms has not yet been developed. Therefore, only by designing a customized chip architecture based on the IoT scenario can we significantly improve performance while reducing power consumption and costs.

Such dedicated chips that can meet the needs of AI computing power and cross-device form factors at the same time are called AIoT chips.

In fact, the AIoT chip is not an independent chip. Currently, the first batch of AIoT chips on the market are almost all composed of core components such as CPU, GPU, FPGA and DSP.

However, there are indeed many differences between AIoT chips and traditional chips.

Here we take the autonomous driving scenario as an example to introduce the differences between AIoT chips and traditional chips:

From a performance perspective, when the autonomous driving encounters a traffic light ahead on the road, if the CPU is used for calculations, the vehicle may have already committed violations such as running a red light, because the calculation speed is too slow and there is no time to react.

If GPU is used for calculation, the calculation speed will be greatly increased, but it will also generate huge power consumption. It is very likely that the car does stop when it encounters a red light, but it may not stop because of the red light, but because all the energy is used for GPU calculation, and there is no extra energy to continue moving forward. In contrast, the use of AIoT chips can not only greatly improve the calculation speed during autonomous driving, but also reduce energy consumption.

From the perspective of process manufacturing, AIoT chips require more specialization in hardware design because the application scenarios and AI algorithms are relatively certain.

Traditional CPU and GPU chips operate using a von Neumann computing architecture and computing model based on instruction streams, while AIoT chips mostly use brain-like or brain-simulated architectures, which can break through the computing bottleneck of the memory wall.

Therefore, compared with traditional chips, AI chips can utilize lower main frequency and smaller chip area to complete AI computing tasks, achieving a perfect balance between diverse needs such as cost, power consumption, and computing power.

These companies have started to layout and seize the market opportunities

Huawei

In addition to its representative HiSilicon Kirin chip in the field of mobile chips, Huawei is also not to be outdone in the field of AIoT chips. The first application field it entered is also very close to people's lives, that is, smart TVs.

The Honghu 818 recently launched by Huawei Honor has very powerful performance. It is equipped with a VideoDecoder module that can support 8K@30Hz video playback. Playing 8K video is equivalent to playing 4 times 4K video. This means that when a device using the Honghu 818 chip opens the current mainstream 4K video, it only takes 1.3 seconds from click to start playing, which is nearly twice as fast as competing products.

Honor has also said that it is making smart TVs with the spirit of making mobile phones. It is conceivable that this product will have many new features that are unimaginable on traditional TVs, and excellent chip performance is the basis for achieving all this.

Baidu

In July 2018, Baidu released China's first cloud-based full-function AI chip "Kunlun" at the AI ​​Developer Conference. It includes the training chip Kunlun 818-300 and the inference chip Kunlun 818-100, which are mainly aimed at AI large-scale computing needs.

One year later, Baidu has delivered another impressive report card. At this year's AI Developer Conference, Baidu launched the Honghu chip, another chip masterpiece following the release of the Kunlun chip.

As an AIoT chip independently developed by Baidu, Honghu is mainly aimed at intelligent voice applications, and is also designed according to automotive standards. It leaves a lot of room for imagination both in smart hardware and in future automotive intelligent voice applications.

Alibaba

In recent years, Alibaba has been particularly active in technological research and development. In addition to the earliest established "Dharma Academy" and "Arhat Hall", it has also established a chip company "Pingtou Ge".

As the saying goes, life and death are indifferent, and if you don't agree, just fight. "Flat-headed Brother" is not behind in the field of chip research and development. Just recently, "Flat-headed Brother" officially released its first AIoT chip, and the name is also very interesting, called Xuantie 910.

You should know that the name "Xuantie" comes from the first magic sword used by Dugu Qiubai in Jin Yong's novel, "heavy sword without edge, great skill without workmanship", which was later melted into Yitian Sword and Tulong Saber by Guo Jing and his wife. Does this mean that Xuantie 910 will become a "magic weapon" in the AIoT chip industry?

Sure enough, according to data, XuanTie 910 has become one of the most powerful RISC-V processors in the industry, with computing power ranking first in the world, and Coremark scores exceeding the second place by more than 40%.

The single-core performance of XuanTie 910 reaches 7.1 Coremark/MHz, and the main frequency reaches 2.5GHz. It has extremely powerful performance and can be applied to 5G, artificial intelligence, autonomous driving and other fields.

In the future, Pingtou Ge will also build an intelligent Internet of Things chip platform for many industries such as automobiles, home appliances, and industry.

Millet

Many people may not imagine that Xiaomi will also be one of the AIoT chip layout. Xiaomi's Huami Technology has released the world's first AI chip in the wearable field, "Huangshan No. 1", which is composed of RISC-V open source instruction set and integrated with AI neural network module. It can perform local data calculation and processing without Internet access. Through the built-in HeartRate, ECGEngine, ECGEnginePro, Arrhythmias four driving engines, Huami Technology has achieved real-time monitoring and analysis of heart rate, electrocardiogram, arrhythmia, etc., which can be widely used in various smart wearable devices.

In April this year, Xiaomi Group issued an internal email announcing that its Pinecone Electronics team was reorganized and part of it was split off to form a new company, Nanjing Dayu Semiconductor, which focuses on IoT chips; Pinecone Electronics will continue to focus on the research and development of mobile phone SoC chips.

It is not difficult to understand why Xiaomi made such a decision. Mobile phone chips require extremely high R&D investment, a long payback period, and long-term technical accumulation.

Since the Surge S1 chip went silent in 2017, Xiaomi's mobile phone chip research and development has been in a deadlock. Even HiSilicon Kirin chips took a long time to go from research and development to commercial use and then compete with international manufacturers.

But in the field of smart home, Xiaomi started its layout very early, and has now formed the advantage of the "Xiaomi ecological chain". AIoT has become Xiaomi's most important strategic direction, and the layout of AIoT chips is also expected.

MediaTek

In April of this year, MediaTek released two AIoT chip platforms, namely i300 and i500.

The i300 is suitable for voice and visual device applications and can be used to develop products for home, IoT and other portable devices; the i500 integrates a more powerful AI processor. The dual-core AI processor (DSP) operates at a frequency of 500MHz, supports deep learning, neural network acceleration and computer vision applications, and is suitable for scenarios such as facial recognition, object recognition, scene recognition analysis, and optical character recognition.

In July, MediaTek launched the latest generation of AIoT chip platform i700, which adopts an octa-core architecture and integrates two ARM Cortex-A75 processors with an operating frequency of 2.2GHz and six Cortex-A55 processors with an operating frequency of 2.0GHz. It is also equipped with an IMG9XM-HP8 graphics processor with an operating frequency of 970MHz and MediaTek's CorePilot technology, and is used in smart cities, smart buildings, smart manufacturing and other fields.

NXP

NXP is one of the hardcore players in AIoT chips. As early as February 2018, NXP launched the AIoT chip, which integrates NXP's i.MX application processor series and Wi-Fi/Bluetooth solutions into a smaller size.

This not only solves the problem of extremely limited size of customized chips in current IoT devices, but also provides higher cross-system efficiency. This forward-looking design also brings a lot of inspiration to later layouters.