At present, several major global trends that drive microelectronics, including 5G, edge computing (in other words, explosive growth of interconnected devices), autonomous driving and software definition cars (SDVs), artificial intelligence (AI), clouds in the cloud Calculation, and high -performance calculation (HPC).
AI, cloud computing, and similar applications such as HPC will never be too much performance. Their bottlenecks are often limited to the space and power consumption of the data center, not the cost of the hardware itself. And even the edge computing device, or the self -driving and software -defined cars, there are many space that can make good use of more computing performance. But in addition to power consumption and space, they will also be limited by more costs (the number of chips and chips that can be used).
AI, car, cloud computing, marginal computing, and HPC. Although these are very different applications, they can be processed through extensive supply of ready -made chips. But on the other hand, a chip tailored for a specific work can achieve performance or functional needs at better cost and power efficiency.
The ARM infrastructure business line product solution department VP Dermot O'Driscoll said: "More and more companies have found that integrated applicable solutions can no longer meet their computing needs. Therefore, they have begun to explore customized SOC projects, Make sure they provide appropriate computing capabilities for the right application. "
In addition, factors such as intellectual property (IP) protection, functional differentiation, design flexibility, and supply stability are more important than ever. This has prompted more and more companies (whether it is an old company or a new company) to develop its own SOC or order directly from the chip foundry.
"Indeed, more and more companies have begun to design chips by themselves." ANSYS technical chief and engineering simulation and EDA software developer Christophe Bianchi said that "the trend of customized chips is particularly obvious in North America, but it can also be seen in EMEA and the Asia -Pacific region. Variety."
"Customized SOC represents an exclusive solution that integrates all your unique differentiated technologies, and it is difficult for others to copy." Graham Curren, founder and CEO of British chip design foundry company Sondrel, said, "It is a chip existing part, Therefore, it is equivalent to protecting the company's intellectual property rights. In addition, because it is tailor -made, it can accurately meet the task needs in terms of power consumption, performance, and volume. "
Apple is a successful case of trying to design its own SOC. Starting from smartphones, developing to personal computers, and finally gained great benefits.
Initially, in the late 2000s, Apple first requested Samsung (Samsung) to introduce some Apple customized technologies in the iPhone processor, and then they quickly established their own chip development team, and began designing highly complicated mobile SOCs. It contains customized CPU microstructures and accelerators. This adds additional value to Apple's iPhone, iPad, Apple TV TV box and other products. In recent years, Apple has begun to create SOC for its personal computer, and has also introduced various customized technologies to further reduce power consumption, extend battery life, and provide higher performance for high -demand applications.
O'Driscoll said: "Several main purpose of designing customized SOC includes an accelerator that introduces specific uses, unique systems or packaging space design, or the optimization cost through vertical integration."
The Apple M series SOC has achieved huge benefits for the company. In the first quarter of 2019, the market share of Apple's personal computer was only 6.4% (about 3.79 million units were sold), but by the first quarter of 2022, it had grown to 9.3% (about 7.32 million units were sold). Apple's R & D expenditure in fiscal year in 2022 was US $ 26.251 billion (US $ 16.217 billion in fiscal 2019), and this huge amount was mainly used in chip development.
In the past, companies with large income such as IBM, Cisco and Sun (now Oracle) have the ability to design their own chips. Apple showed the potential of consumer electronics SOC to the world, and companies such as Huawei and Samsung also followed.
By the mid -2010s, the scale and profit of the Internet giants were huge enough to design their own chips. Today, auto manufacturers have also begun to embark on Apple's route.
"In recent years, the number of companies that have designed chips has grown sharply." Dan Hutchenson, deputy president of Techinsights, pointed out, "Originally, this approach was regarded as the result of the natural evolution of the trend of self -produced ASIC in the old age. Began to use the production model of an uncoded factory/outsourced foundry. But the involvement of Apple quickly subverted the industry of the entire mobile application processor. By 2015, Google, Meta, Amazon, Baidu (BAIDU) Companies such as Tesla have begun designing their own chips. Subsequently, venturers have begun to fund new companies in AI and new energy. Since 2021, major car manufacturers in the world have also begun designing their own chips. ","
Hyperscalers lead the industry
At present, Apple is the only large -scale personal computer manufacturer who designed its own processor. Compared with Apple's complete control over its Macos, general computer manufacturers have no significant influence on Windows development. When their success must depend on the support of Microsoft, it is naturally lacking in interest in the development of SOC.
Different from personal computer manufacturers, the large -scale data center enterprise (Hyperscale) has a complete control of hardware and software, so it is very suitable for developing its own exclusive SOC.
Bianchi said: "The biggest promotion factor in the initial period of customized chips is the huge demand of AI and cloud computing from oversized large -scale data centers."
Google provides many most popular network services in the world, and uses a diversified customized chip strategy to promote its data center, including the use of tension processing unit (TPU) to accelerate AI training and inference, as well as using image transcoding units (image transcoding unit (image transcoding unit (image transcoding unit (image transcoding unit (image transcoding unit ( VCU) process YouTube pictures. According to reports, they are also developing server -specific SOCs. Whether the server SOC in this rumor will be more suitable for Google than AMD's Epyc or Intel's Xeon CPU, it remains to be observed.
In any case, TPU and VCU have brought Google to expand the performance, functions, ductility, and cost benefits required for expanding AI and picture services.
Karl Freund, the founder and chief analyst of the Industrial Analysis Company Cambrian-AI Research, said: "The demand for cost and customization promotes these companies to develop its own chips." He believes that AI computing is that these super-large-scale data centers want to via their own chips through their own chips. One of the most important tasks for treatment.
"Super -large -scale data center companies make a chip by themselves than buying a lot cheaper, and save the cost of middlemen." Users of their own cloud services, and have high -value specialized software. They can create exclusive hardware with better energy efficiency for these software. "
Amazon's "Amazon Web Services" (AWS) is one of the world's top cloud services. It not only promotes Amazon's own online stores and warehousing management systems, but also uses various network services such as Netflix. But AWS's chip strategy is different from Google.
In order to provide customers with sufficient functions and computing performance on the premature predictable power consumption, AWS designed a server -level Graviton SOC for its data centers for its data center. AWS's Graviton SoC has now developed to the third generation.
John Koeter, Vice President of Marketing and Strategy of the SynopSys solution department, said: "Obviously more and more companies (especially large -scale data centers) have begun to choose to design their own chips, and to meet differentiated work needs in the data center. "
In companies like AWS, Baidu, Google, Meta, or Microsoft Azure, a large number of diversified workloads are operated. In order to further improve the efficiency, these tasks will be transformed into operation through customized chips sooner or later, and these companies all have the strong financial resources required for developing customized SOCs.
"We see the biggest growth from the field of data infrastructure, including clouds, data centers, networks, storage devices, and 5G infrastructure." Sudhir Mallya, director of marketing of Alphawave SEMI, said: "From today, customized, customized The growth rate of the chip in the infrastructure of data center is really amazing. From a few years ago, oversized data center companies such as Google, Microsoft, AWS, Facebook (META) have begun to design their own chips, and we have seen this trend. "
AI and HPC: Different strategies
At the same time, many manufacturers' data center -level SOCs are developed specifically for AI and HPC applications. Among them, both processors, SOCs, or multi -chip system -level packaging (SIP) are customized. Koeter said that these customized AI and HPC solutions can often provide very powerful performance, but manufacturing is also very difficult and expensive, because some chips are as large as the size of an EUV mask (858mm 2).
Older manufacturers provide a lot of standardized products suitable for AI and HPC applications. For example, AMD has Instinct computing GPU and EPYC processor (using ChiPlet) to solve AI and HPC workloads.
Intel uses a multi -directional strategy, using the single Habana processor to process AI applications, multi -chip's Data Center GPU MAX processing AI and HPC applications, and the fourth -generation XEON can expand the other applications of multi -chip.
In contrast, NVIDIA's latest H100 products are designed by monolithic. These are ready -made products designed by companies with huge R & D budgets and talents, and have their own unique advantages. Nevertheless, many new players are still going to enter the market in the next few years to challenge these industry giants.
For more electronic components requirements,pls see:https://www.megasourceel.com/products.html