Huawei Chip 20 Years

Source: Content from " Semiconductor Weathervane ", author: Electronics Chief Chen Hang, thank you.

Huawei established its ASIC design center in 1991, then established HiSilicon Semiconductor in 2004, and has become the representative of China's independent chip design. So what chips has Huawei made?

From a broad perspective, Huawei mainly designs five types of chips:

1. SoC chip

2. AI chips

3. Server chips

4. 5G communication chips

5. Other special chips

Huawei Chip Panorama

1. SoC chip (Kirin series): Mobile phone SoC chips have always been Huawei's main research focus. With the Kirin 980 processor launched on August 31, 2018 and the Kirin 985 chip expected to be launched in the second half of this year, Huawei's mobile phone chips have reached world-class levels.

2. AI chips (Ascend series): On October 10, 2018, Huawei released two AI chips, Ascend 910 and Ascend 310, at its HC conference. They are manufactured using 7nm and 12nm process respectively. The Ascend series AI chips use Huawei's pioneering unified and scalable architecture, the "Da Vinci architecture", which covers all scenarios from extreme low power consumption to extreme high computing power.

3. Server chips (Kunpeng series): Huawei optimized and adjusted the design of the technology licensed by its partner ARM, and released the Kunpeng 920 and the Taishan server and Huawei cloud service based on Kunpeng 920 on January 7, 2019.

4. 5G communication chips (Balong and Tiangang series): Huawei's 5G chips are mainly divided into terminal chips (Balong series) and base station chips (Tiangang series). The Balong series is a mobile terminal baseband chip and has always been a dedicated chip for Huawei mobile phones. On January 24, 2019, Huawei launched the industry's first 5G-oriented base station core chip (Tiangang chip) and 5G multi-mode terminal chip (Balong 5000).

5. Other special chips: (router chips, NB-IoT chips, IPC video codec and image signal processing chips, etc.): Lingxiao series are mainly used for home access products; Lierda NB-IoT module is the world's leading narrowband Internet of Things wireless communication module; IPCSoC chip covers the core technologies of video surveillance - ISP technology and video codec technology.

The core of all mobile phone chip solutions are two parts: AP and BP. AP stands for Application Processor, which includes CPU (Central Processing Unit) and GPU (Graphics Processing Unit), and BP stands for Baseband Processor, which is responsible for processing various communication protocols, such as GSM, 3/4/5G, etc. In addition, there are core units such as RF, which together constitute the chip of SoC chip integrated circuit.

SOC can effectively reduce the development cost of electronic/information system products, shorten the development cycle, and improve product competitiveness. It is the main product development method that the industry will adopt in the future. Huawei's SOC chip is designed and developed by HiSilicon Semiconductor.

Huawei HiSilicon chip SOC structure

The development of HiSilicon chips has gone through a long period of time.

In 2004, Huawei HiSilicon began to develop mobile phone chips, and in 2009, it launched a GSM low-end smartphone solution as scheduled, using the Windows mobile operating system. The BP technology in this solution is self-developed, and the technology comes from Huawei's GSM base station. The AP chip is called K3V1, and the process is 110nm, which is behind the 65/55/45 nm process used by competitors at the time. In addition, the operating system chosen was the declining Windows Mobile, so the performance of the first generation of HiSilicon chips was not ideal and was quickly eliminated by the market.

In 2012, Huawei released K3V2, claiming to be the world's smallest quad-core ARM A9 processor. The most fundamental difference from K3V1 is that K3V2 uses ARM architecture and supports Android operating system instead of Windows Mobile. At the same time, Qualcomm APQ8064 and Samsung Exynos4412 have already used 28 and 32nm processes. K3V2 uses 40nm process, generates a lot of heat, has poor game compatibility, and has not been recognized by the market.

In May 2015, Kirin 650, the world's first low-end chip manufactured using the 16nm FinFET Plus process, was released, leading Honor 5C and G9 to continue to exceed 10 million sales.

On August 20, 2015, Kirin chip shipments exceeded 100 million, with an average daily shipment of 290,000 chips within two months.

In November 2015, Huawei released the Kirin 950 SoC chip, which once again ranked first in comprehensive performance. With its performance and process advantages, it successfully led Qualcomm by half a year. This chip has been used in Huawei's Mate 8, Honor 8, Honor V8 operator customized version and standard full network version.

In April 2016, the Kirin 955 SoC chip was released, which upgraded the A72 architecture from 2.3GHz to 2.5GHz, leading the P9 series to become Huawei's first flagship phone with sales exceeding 10 million.

On September 2, 2017, Huawei released the AI ​​chip Kirin 970 and used it in Huawei Mate 10 , which was officially released in Munich, Germany on October 16, 2017. So far, Huawei has successfully released the latest new generation successor, the second-generation AI chip Kirin 980, which has once again raised the level of flagship mobile phone CPU to a new height.

The historical evolution of Huawei HiSilicon chips

The major mobile phone chip manufacturers on the market are Apple, Huawei, Qualcomm and Samsung.

Apple A series: Apple's A12 processor is the world's first 7-nanometer chip, and its performance ranks first. However, since Apple does not have its own baseband technology, GPS/WIFI chips need to be purchased from outside, so the A series chips never contain baseband parts and do not need to assume GPS/WIFI functions;

Huawei Kirin Series: Huawei's Kirin 980 adopts the most advanced eight-core design with a maximum main frequency of up to 2.8GHz. It is expected that Kirin 990 will also inherit the 5G baseband to achieve true 5G full network access;

Qualcomm Snapdragon series: The Snapdragon 855 processor released, which uses the new Kryo 485 architecture and 7nm process technology, improves graphics rendering by 20%, CPU performance by 45%, and has a maximum main frequency of 2.84GHz, which is also very powerful;

Samsung Exynos series: The Exynos 9810 processor is based on Samsung's self-developed M3 architecture, with 4 2.9GHz M3 large cores and 4 1.9GHz A55 small cores, and a 10nm manufacturing process.

Comparison of major mobile phone chip manufacturers around the world

A server is a high-performance computer that acts as a node in a network and stores and processes 80% of the data and information on the network. Therefore, it is also called the soul of the network. The core part of a server is the server chip, which can be said to be the brain of the entire server. However, it is precisely because of its importance that its technical difficulty cannot be underestimated.

On January 7, 2019, Huawei released the Kunpeng 920 server chip. The chip is developed based on the ARMv8 instruction set, uses a 7nm process, has up to 64 cores, supports 8-channel DDR4 memory and PCIe 4.0 protocol, and therefore has its own unique performance advantages: 1. The performance score exceeds the previous industry benchmark product by 25%, and the energy efficiency is improved by 30%, but the power consumption is reduced; 2. The dual-port design makes the speed of this chip reach twice that of mainstream products in the industry; 3. The memory bandwidth of Kunpeng 920 exceeds the mainstream industry by 46%, and the network bandwidth is increased by four times. Huawei Kunpeng chips have now reached the industry-leading level in performance.

The server chip market is a lucrative market. Previously, server chip technology has been monopolized by the United States as the core of technology. According to DRAMeXchange data, 97% of the world's server processors are X86 architecture, so Intel is the absolute overlord in the server field. The launch of the Kunpeng 920 chip also plays an extremely critical role in the leap in the level of domestic servers. Although Huawei announced that the Kunpeng 920 processor is only for its own use, referring to the development history of the Kirin series processors, we have reason to believe that the ARM architecture processor can be developed and strengthened by self-purchased and self-used Kunpeng 920 processors. Considering that the server market is mainly for enterprise users, this process will be more difficult and take longer than that faced by mobile phone processors.

Huawei server chip performance comparison with other manufacturers

In addition to the Kunpeng 920 processor, Huawei also launched three TaiShan series servers, which use Kunpeng 920, including TaiShan 22080, Taishan 5280/5290, and TaiShan X6000, which are respectively aimed at the balanced server, storage server and high-density server markets.

Currently, chip development is beginning to face a paradoxical dilemma: on the one hand, processor performance is facing physical limits and cannot grow according to Moore's Law; on the other hand, with the rise of new applications such as cloud, deep learning, and AI, the volume of data has exceeded the growth rate of "Moore's Law" in terms of computing performance requirements. The natural development of processors cannot meet the needs of high-performance applications, and a technology gap has emerged. In this case, the use of heterogeneous computing methods with dedicated coprocessors to improve processing performance has become the most ideal and convenient solution.

Heterogeneous chips usually have four architecture options in applications: CPU, GPU, FPGA, and ASIC. They can provide specialized hardware acceleration to implement key processing functions required in various applications, which is sufficient to meet the personalized needs of chips under the AIoT platform. CPU and GPU are common general-purpose chips. CPU is suitable for logic control, serial operations and general-purpose data operations. GPU has a large-scale parallel computing architecture and is good at handling multiple tasks such as graphics computing. Since deep learning usually requires a large number of but not complex training algorithms, GPU is more suitable for deep learning operations than CPU. FPGA (field programmable gate array) is a semi-custom chip that has been programmable, has the advantage of parallel processing, and can also be designed with multiple cores. The biggest advantage of FPGA is its programmable nature. Users can quickly burn the circuit according to the required logical functions to realize customized hardware functions.

Compared with general-purpose chips such as CPU, GPU and semi-customized FPGA, the computing power and efficiency of ASIC chips are directly customized according to the needs of specific algorithms. Therefore, ASIC chips have the advantages of small size, low power consumption, high reliability, high computing performance and high computing efficiency. Therefore, in the specific application fields it targets, the energy efficiency of ASIC chips is far superior to general-purpose chips such as CPU, GPU and semi-customized FPGA.

Four heterogeneous chip options in the AI ​​era

In 2018, Huawei released two cloud data center AI chips: Ascend 910, which has the highest single-chip computing density, and Ascend 310, an AI SoC with extremely efficient computing and low power consumption .

Ascend 910 belongs to the Ascend Max series and is used in data center servers. Its performance exceeds Nvidia's most powerful chip AI V100 and is the AI ​​chip with the highest single-chip computing density released in the world. Based on Ascend 910, Huawei will also launch a large-scale distributed training system Ascend Cluster, which links 1024 Ascend 910 chips to form an AI computing cluster, providing ultra-high-level AI computing capabilities with a maximum computing power of 256P, bringing AI training speed to a new level.

Comparison of computing power between Ascend 910 and Ascend cluster and existing products

The Ascend series of AI chips uses Huawei's pioneering unified and scalable architecture, the "Da Vinci Architecture" , which achieves full coverage from extreme low power consumption to extreme high computing power scenarios . Currently, no other architecture in the global market can do this.

The "Da Vinci Architecture" can develop deployment, migration, and collaboration suitable for all scenarios at one time, greatly improving the efficiency of software development and accelerating the application of AI in various industries.

Huawei highlighted the full-stack, full-scenario AI solution at the 2018 All Connected Conference. The full-stack, full-scenario AI solution is an integration of Huawei's previously released Huawei Cloud EI for governments and enterprises, and HiAI for smart terminals.

The vertical axis full stack is a technical functional perspective, which refers to the full stack solution including chips , chip enablement , training and inference frameworks , and application enablement .

The horizontal axis of the full scene includes consumer terminals, public cloud, private cloud, various edge computing, IoT industry terminals these five categories of scenes.

Huawei's full-stack, all-scenario AI solution

5G refers to the fifth generation of mobile communication technology, which is an extension of 4G. Its peak theoretical transmission speed can reach tens of Gb per second, which is hundreds of times faster than the network transmission speed of 4G. At the same time, 5G also has millisecond-level transmission delay and hundreds of billions of connection capabilities. It is the communication foundation for the interconnection of all things and deep interaction between man and machine.

Unique advantages of 5G over 4G

The underlying technology of 5G communication is 5G chip, which is mainly divided into RF chip and baseband chip .

RF chips are mainly used to receive and transmit signals, including PA (power amplifier), filters, switch chips, etc. Baseband chips are mainly used to modulate and demodulate transmitted signals. As a global leader in communications equipment, Huawei has a large number of patent reserves, so it has strong baseband chip design capabilities. In contrast, although Apple's A series chips have reached world-class levels, they lack baseband chip design capabilities, so they need to use third-party baseband chips.

Compared with 2G, 3G, and 4G baseband chips, 5G baseband chips have many design difficulties. 1. Design complexity caused by multi-band compatibility. 2. Supporting millimeter waves has also become a design difficulty for 5G baseband chips. 3. The increase in the number of supported modes has also increased the design difficulty. 5G baseband chips need to be compatible with 2G/3G/4G networks at the same time.

At the pre-communication meeting of the 2019 Mobile World Congress, Huawei released two 5G chips, namely the 5G base station core chip (Huawei Tiangang) and the baseband chip for 5G terminals (Balong 5000).

Tiangang chip - the industry's first 5G base station core chip. Huawei Tiangang is the world's first chip with super integration, super computing power, and ultra-wide spectrum. It has brought revolutionary improvements to AAU, reducing the size of base stations by more than 55%, reducing weight by 23%, saving power by 21%, and saving half the installation time compared to standard 4G base stations, effectively solving challenges such as difficult site acquisition and high costs. At the same time, the Tiangang chip allows most base stations on the market to be directly upgraded to 5G, which means that 4G to 5G can be directly upgraded without changing the power supply or powering off, greatly facilitating the update of 5G.

Huawei 5G Tiangang chip application and enabling concept diagram

Balong 5000 - the world's first commercial chipset that supports 3GPP standards for 5G. Balong 5000 has five world records, one of which is world-leading: the world's leading multi-mode single-core module integrating 2G, 3G, and 4G; the world's fastest speed, @Sub-6 GHz 200MHz: downlink speed 4.6Gbps, uplink speed 2.5Gbps; the world's first uplink/downlink decoupled multi-mode terminal chip; the world's first chipset that supports both NSA and SA architectures; the world's fastest peak downlink speed @millimeter wave 800MHz Gbps; the world's first R14 V2X on a 5G chip.

Balong 5000 supports 5G across all frequency bands, providing a complete 5G solution. It is the key to fully opening the 5G era. It can support a variety of rich product forms, including home broadband terminals, car terminals and 5G modules in addition to smartphones, and will bring consumers a different 5G connection experience in more usage scenarios.

5G terminals of mainstream manufacturers: Commercial CPE in 2018, Commercial smartphones in 2019

Huawei has released the first 5G terminal product based on the Balong 5000 chip: 5G CPE Pro. This is the world's fastest 5G CPE, supporting Wi-Fi 6 technology. The main application scenario is smart home. 5G CPE Pro supports 4G and 5G dual-mode. Under the 5G network, it can download 1GB of high-definition video in 3 seconds, and even 8K video can be opened in seconds without lag, setting a new network speed standard for small CPE.

Huawei 5G CPE Pro basic information