AI chips are in short supply, how can large manufacturers break the situation?
The arrival of the wave of artificial intelligence has given rise to a huge demand for computing power. In the past, computing power was mainly used in scientific computing, big data and other fields, but now the explosive growth of AI applications has made the demand for computing power rise exponentially.
Large language models, such as ChatGPT, need to consume a lot of computing power for training and reasoning. It is estimated that the computing power required to train ChatGPT is equivalent to the carbon emissions of 3 human lifetimes. This huge demand for computing power has made the supply of AI chips in short supply, which has become a bottleneck restricting the development of AI.
Faced with this challenge, tech giants have taken action. Companies such as Google, Meta, Amazon, and Microsoft are ramping up efforts to develop their own AI chips to reduce their dependence on external suppliers such as Nvidia. They hope to reduce the cost of AI applications by customizing the design to improve chip performance and energy efficiency.
Self-developed AI chips are not an easy task. It requires strong capital investment, technology accumulation and talent reserves. For most companies, it will be difficult to completely get rid of their dependence on existing chip suppliers in the short term. They are also looking for outsourcing cooperation to speed up the process of self-development by using the services of professional chip design companies.
Whether it is independent R&D or outsourcing cooperation, tech giants are fully prepared for the battle for AI computing power. Because whoever can control the computing power can lead the development direction of AI and gain huge business value and strategic advantages.
Status quo and challenges of AI chips
The current AI chip market pattern
In the current AI chip market, NVIDIA occupies a monopoly position with its strong GPU product line and mature CUDA ecosystem. According to industry data, in the fourth quarter of 2021, NVIDIA's global GPU computing power chip market share was as high as 95.7%.
In addition to NVIDIA, traditional chip manufacturers such as AMD and Intel are also increasing investment in AI chips. AMD's Instinct series of GPUs has entered the AI accelerator market, while Intel has launched dedicated product lines such as Gaudi and Habana, which are AI chips.
Domestic chip companies are also stepping up on the AI track. AI chip products from Haiguang Information, Cambrian, Tiantian Zhixin and other companies have been launched and have gained a certain market share. Although these domestic chips cannot compete with NVIDIA's top products in terms of computing power for the time being, they have shown excellent performance in specific scenarios.
The root cause of the chip shortage
The reasons for the current tight supply of AI chips are mainly due to the following aspects:
1. The surge in demand has exceeded expectations. With the explosion of AI applications, the demand for computing power far exceeds the expectations and capacity planning of chip manufacturers.
2. Manufacturing capacity can't keep up. The chip production capacity of advanced processes has always been a limiting factor, and AI chips have extremely high requirements for process technology, which makes the capacity bottleneck more prominent.
3. Geopolitics affects supply chains. The export controls imposed by the United States on Huawei, Nvidia and other companies have also exacerbated the supply shortage of AI chips to a certain extent.
4. Ecological construction is lagging behind. AI chips require not only hardware strength, but also a well-established software ecosystem. The shortcomings of domestic chips in this regard also restricts their pace of expansion in the market.
In the face of these challenges, chip manufacturers are increasing investment, expanding production capacity and accelerating the pace of new product launches. However, in the short term, the contradiction between supply and demand of AI chips will still persist.
The response of large manufacturers
Self-developed customized AI chips
In order to gain greater initiative and reduce dependence on external suppliers, tech giants have launched plans to develop their own AI chips.
Google is one of the best of them. The company's self-developed TPU (Tensor Processing Unit chip) has been widely used in Google's AI services. The latest generation of TPU v4 chips, with a single-precision floating-point operation capability of up to 92 trillion times per second, surpasses the performance of NVIDIA GPUs of the same period in specific tasks.
Amazon also launched its self-developed Graviton series of ARM chips in 2021 to support its cloud servers. The company also acquired Israeli AI chip startup Annapurna Labs to accelerate the development of AI chips.
As for Microsoft and Meta (Facebook), they have adopted more of an investment and outsourcing strategy. Microsoft invested in OpenAI to gain the right of first refusal to use TPU; Meta, on the other hand, has established in-depth cooperative relationships with companies such as Nvidia.
Strategic investment layout
In addition to independent research and development, technology giants are also laying out the AI chip industry chain through investment and acquisition.
Google has not only invested in Anthropic, an AI company, but has also acquired a number of startups focused on chip design, such as Tensor Machinery.
Nvidia acquired Arm in 2022 to expand its presence in the data center and AI chip markets.
As for Apple, Amazon, and Meta, they also have their own investment maps, covering multiple fields from chip design to AI algorithms.
These investments and mergers and acquisitions will not only help expand the company's technological capabilities, but also prepare for the future AI chip ecosystem.
Open source ecosystem construction
The ecosystem has always been an important reason for companies like Nvidia to achieve a monopoly position. In order to get a share of the AI chip market, domestic and foreign manufacturers are working hard to promote the construction of open source ecology.
For example, both Google and Meta are promoting the RISC-V open-source instruction set architecture. RISC-V is considered to be an important direction for future chip design, which is expected to help break the monopoly of the existing ecosystem and attract more developers to join.
For another example, some domestic chip companies are independently developing a CUDA-compatible software ecosystem to reduce the migration cost of users.
Open-source AI frameworks such as TensorFlow and PyTorch are also expected to play an important role in the future AI chip ecosystem.
The construction of an open source ecosystem will help lower the barrier to entry and promote the diversified development of the AI chip industry.
foreground
The pace of domestic substitution has accelerated
Under the influence of geopolitics, the mainland government and enterprises are increasing their support for domestic AI chips. On the one hand, policy dividends such as tax incentives and special fund support have been given at the policy level; On the other hand, leading enterprises are also increasing investment in independent research and development and industrial layout.
Judging from the current development trend, a number of domestic AI chip companies, such as Haiguang Information and Cambrian, have made breakthroughs, and their product performance is expected to be equal to or even surpass similar foreign products in specific scenarios.
Domestic chip design companies such as VeriSilicon are also growing rapidly in the tide of AI chips, and their IP cores and EDA tools provide important support for domestic chips.
Compared with foreign giants, there is still a certain gap between domestic AI chips in terms of computing power level, process technology, and ecological construction, but this gap is gradually narrowing. Driven by independent and controllable demand, the pace of domestic substitution will be accelerated.
Ecological diversification
The AI chip ecosystem will show a trend of diversified development. On the one hand, general-purpose chips and custom chips are developing in parallel; On the other hand, hardware and software will enable deeper convergence and innovation.
In the field of general-purpose chips, traditional manufacturers such as Nvidia, AMD and Intel still dominate. But they are also ramping up the research and development of customized chips to achieve higher performance and energy efficiency.
Customized AI chips launched by Google, Amazon and other companies will play a unique role in specific scenarios.
From the perspective of software, the rise of the open source ecosystem will promote innovation in AI frameworks and programming models, thereby providing strong support for hardware innovation. Ecosystems such as CUDA and RISC-V will play an important role in the future.
The AI chip ecosystem is undergoing a process of diversification. The integration and innovation of hardware and software will inject new impetus into the development of AI.
AI computing power competition holds
The demand for computing power for large model training is becoming a new driving force. According to industry forecasts, the global data center AI acceleration chip market will reach $400 billion in 2027, which will further intensify the battle for AI computing power.
In this race, Nvidia still has the advantage. The company's latest H100 and Blackwell series chips will further expand its leading position in the AI accelerator market.
AMD, Intel and domestic chip manufacturers are also catching up. They have launched a new generation of AI accelerator products, which are expected to compete with NVIDIA in specific scenarios.
Customized AI chips developed by Google, Amazon and other companies will also add new variables to this competition.
In the next few years, the battle for AI computing power will become fierce. Chip performance, energy efficiency, and cost will become the focus of competition, and the construction of the ecosystem will also determine the final victory.