news

ai chips are also called ai accelerators or computing cards. they can be divided into general-purpose graphics processing units (gpgpu), central processing units (cpu), application-specific integrated circuits (asic), and field-programmable gate arrays (fpga) according to their technical architecture. in the development of artificial intelligence, researchers have found that parallel computing can be used for efficient model training and processing of large-scale complex data. gpgpu has higher parallel computing performance than other chips and is suitable for computing-intensive applications, so it has become the mainstream of computing chips.

according to functional classification, ai chips can be divided into two types: training cards and inference cards. training cards, also called large cards, usually have higher computing power and memory bandwidth to support large amounts of computing and data processing during training; inference cards, also called small cards, have lower parameters and only need to meet inference requirements. in general, training cards can be used as inference cards, but inference cards cannot be used as training cards. simply put, the training of large models requires a large number of training cards to form a graphics card cluster, while in applications, inference cards are required to run ai models for calculations.

this round of artificial intelligence wave was set off by chatgpt, and took language large models and generative ai applications as the entry point. since google published it in 2017, transformer has not only brought c-end hit products such as chatgpt, but has also been widely used in natural language processing, computer vision, autonomous driving and other fields. various chinese and foreign technology companies continue to increase their investment in related fields, including google, meta, microsoft, bytedance, baidu and other domestic and foreign technology giants and start-ups all hope to get a piece of the pie, and other non-technical companies are also constantly making arrangements in terms of talent, technology and resources. according to bloomberg intelligence's forecast data, by 2032, the proportion of generative ai in total information technology hardware, software, services, advertising and games expenditures may expand from less than 1% at present to 12%.

in 2018, openai launched the first generation of generative pre-trained language model gpt-1, which has 117 million parameters. the gpt-4 released in 2023 has about 1.8 trillion parameters. the increase of more than 10,000 times in the number of parameters in 5 years has brought about an exponential increase in the demand for computing power for large models. openai estimates that since 2012, the computing power demand for training the world's top ai models has doubled every 3-4 months. especially after 2023, ai is developing in full swing, and manufacturers at home and abroad have increased their investment in generative ai. the "hundred-model war" has re-emerged, which has also caused the demand for computing power to be in short supply for a long time.

domestic ai chips are mainly focused on the inference side, which is tonvidiaon the other hand, with the development and popularization of ai technology, the industry's demand for high-performance, low-power ai chips is increasing. however, there are also manufacturers who are planning to develop training-side chips. at the 2024 global ai chip summit, biren technology announced its original heterogeneous gpu collaborative training solution hgct for the first time. for the first time in the industry, it supports three or more heterogeneous gpus for mixed training of the same large model. moore thread announced its mtt s4000-based wanka intelligent cluster kua'e (kuae), focusing on training and testing. huawei, cambrian, haiguang information and other manufacturers have also taken similar actions.

ai chip startups are extremely difficult. laimi research institute believes that the first difficulty in ai chip startups is to solve chip design and manufacturing problems. due to overseas restrictions, related design tools have been regulated. however, manufacturing issues are more realistic problems. nvidia's latest ai chip uses a 4nm process. currently, only tsmc has the manufacturing capacity, while mainland china's chip manufacturing capacity is slightly insufficient.

the second is the market issue. nvidia is still the most cost-effective ai chip solution. domestic ai companies purchase domestic ai chips for subjective or objective reasons of domestic substitution, and on the other hand, they also hope to use products that are more in line with their needs (such as bytedance and other manufacturers designing ai chips based on their needs). on the one hand, the big model has accelerated the development of ai chips, but on the other hand, it has also widened the gap between chip manufacturers. domestic ai chip manufacturers must consider what specific needs their products can meet for customers in order to make customers willing to pay a high premium.

the third is profitability. for well-known reasons, the development cycle of ai chips is long, intensive, and costly, and these investments are often difficult to get returns in the short term. therefore, it is extremely important to be able to generate revenue. under the current dynamics of the primary market, it is unlikely to rely entirely on the venture capital market for transfusions, and the positive cash flow of ai chip companies is also worth testing.

as the big model revolution sweeps the world, computing power demand has reached a new high, driving the iteration and evolution of cloud, edge, and side ai chips. under the three clouds of explosive data growth, technology approaching physical limits, and complex and changing international situations, many ai chip companies are moving forward under pressure in a low-key and pragmatic manner, actively preparing for the opportunities brought by the generative ai wave.

as elon musk said: the only limitation is raw materials, and every link in manufacturing can be completely reinvented. the huge potential of ai will inevitably give rise to a market size of trillions of dollars. the "bottleneck" is only temporary. with the breakthrough of advanced processes and capital investment, the distance between us and overseas ai is shrinking rather than increasing. the space for domestic substitution is huge, and related companies will also usher in opportunities for rapid growth.

endgame thinking requires locking in the winner. the future world needs infinite computing power, but referring to the competition pattern of the us market, there may be only a few winners. due to the winner-takes-all nature of ai chips, we believe that domestic ai chip companies must have technical strength, channel capabilities, and financing capabilities to win. these capabilities pose huge challenges to startups, but they also build industry barriers invisibly.