The ChatGPT boom has somewhat eased the downbeat atmosphere in the semiconductor industry, with recent market news indicating that TSMC’s 5nm demand has suddenly surged and capacity utilisation may be at full capacity in the second quarter. Semiconductor supply chain industry insiders revealed that TSMC’s urgent orders come from Nvidia, AMD and Apple’s AI and data centre platforms, and ChatGPT’s burst of fire has increased customer pulling momentum.
Recently, a chatbot called “ChatGPT” has been making a name for itself in the semiconductor world. Launched by AI lab Open AI at the end of November last year, the AI program has registered over one million people within five days of its launch, and by the end of January this year it had surpassed 100 million.
Officially, ChatGPT’s “Chat” refers to chat, which is its presentation format; “GPT” stands for Generative Pre-trained Transformer, which is the pre-trained model that supports its operation. The “GPT” stands for Generative Pre-trained Transformer, the pre-trained model that supports its operation.
In general, ChatGPT is an ultra-intelligent conversational AI product based on a large-scale language model. Whether we are discussing AIGC (AI Generated ContentAI Produced Content), the latest form of content creation on the Internet, or the explosive ChatGPT, the essence is to explore the AI industry chain behind it.
Part.1
AI chips: GPU, FPGA and ASIC will benefit
The three major elements of artificial intelligence are data, algorithms and arithmetic power. ChatGPT, which is based on OpenAI’s third-generation model GPT-3, is the ultimate source of arithmetic power, and ChatGPT’s explosion represents a new round of breakthrough in AI chip technology. According to public information, AI arithmetic chips refer to accelerated AI applications and are mainly divided into GPU, FPGA and ASIC.
GPU
CPUs are generally used with acceleration chips because their arithmetic power is very limited and they struggle to handle parallel operations. In the AI era, GPU occupies a larger share of the cloud-based training chips and is regarded as the core of arithmetic power in the AI era. In the GPU market landscape, Nvidia, AMD and Intel have almost monopolised the entire GPU industry in terms of revenue.
Currently, the computing cluster behind ChatGPT uses Nvidia’s AI chips. openAI has said that ChatGPT is a super AI done in collaboration with Nvidia and Microsoft. microsoft built the super computer cluster in its own cloud, Azue HPC Cloud, and made it available to The supercomputer is reported to have 285,000 CPU cores and more than 10,000 AI chips.
Although Nvidia has the first mover advantage in this case, there are many other companies in the market catching up, such as Google’s Tensor Processor TPU, Baidu’s Kunlun series and Alibaba’s Contained Light 800.
FPGAs
FPGA (Field Programmable Gate Array), also known as Field Programmable Gate Array, is a digital integrated circuit that changes and configures the internal connection structure and logic units of the device through software means to complete the established design functions.
FPGA chips have obvious advantages in terms of real-time (fast data signal processing speed) and flexibility, and can also be programmed and computed in parallel, occupying an irreplaceable position in the field of deep learning.
With higher speed and very low computational energy consumption compared to CPUs/GPUs/ASICs, FPGAs are often used as a low-volume alternative to dedicated chips. To implement deep learning functions when building AI models, FPGAs need to be combined with CPUs and applied together to deep learning models, which can also achieve huge arithmetic power requirements.
In terms of market pattern, the global FPGA chip market is dominated by two companies, Ceres and Intel, which occupy the majority of the market share. As FPGA chips have high technical and financial barriers, Chinese companies have a large gap in this field. In recent years, China’s leading enterprises have also achieved some technological breakthroughs in FPGA chip chips.
ASIC
ASIC (Application Specific Integrated Circuit), or special-purpose integrated circuit, whose computing power and computing efficiency can be customized according to specific user needs, is widely used in artificial intelligence devices, virtual currency mining equipment, consumables printing equipment, military defense equipment and other intelligent terminals.
ASIC chips can be divided into TPU chips, DPU chips and NPU chips depending on the function of the terminal. Among them, TPU (Tensor Processing Unit) is a tensor processor, dedicated to machine learning. DPU (Data Processing Unit), can provide engines for computing scenarios such as data centres. simulates human neurons and synapses and uses a deep learning instruction set to directly process large-scale electronic neuron and synapse data.
Compared to GPUs and FPGAs, ASICs lack flexibility, especially in such fields as AI and servers, where the features of ASIC chips become a liability in the face of constant iteration of various algorithms. However, Horizon CEO Yu Kai has publicly stated that once software algorithms are fixed, dedicated integrated circuits ASICs must be the future direction, and on a per-watt power consumption computing power basis, ASICs can be 30-50 times more powerful than GPUs, which will be the focus of future industry competition.
At present, Google, Intel, Nvidia and other technology giants have released ASIC chips such as TPU and DPU, and major domestic manufacturers have also started to target this market quickly, for example, Cambium has launched a series of ASIC acceleration chips, and Huawei has also designed the Rise 310 and Rise 910 series of ASIC chips.
Part.2
HBM/Chiplet expected to benefit
Overall, with AIGC (AI Generated Content) driving AI industrialization from software to hardware switching, the semiconductor + AI ecosystem is gradually becoming clearer and AI chip products will be implemented on a large scale. The core of the hardware side includes AI chips/GPU/CPU/FPGA/AISoC, etc. In AI chips, arithmetic power and information transmission rate become key technologies, and the balance of chip performance and cost also drives the surrounding ecology, including HBM/Chiplet and other industry chains to benefit.
01
Emerging storage HBM
Publicly available information shows that AI conversation programs require high-capacity, high-speed storage support during the execution of calculations, and the industry expects that AI chip development will also further expand the demand for high-performance storage chips.
Samsung Electronics, for example, has indicated that demand for high-performance high-bandwidth memory (HBM), which provides data for GPUs and AI accelerators, will expand. In the long term, demand for high-performance HBM of 128GB or more for CPUs and high-capacity server DRAM is expected to increase as AI chatbot services expand.
Recently, Korean media reported a rapid increase in HBM orders from two storage majors, Samsung and SK Hynix, after the start of 2023, and prices have gone up, with market sources revealing a recent five-fold increase in HBM3 specification DRAM prices.
02
Chiplet
Chiplet heterogeneous technology can not only break through the blockage of advanced processes, but also significantly improve the yield of large chips, reduce the complexity of design and design costs, and reduce chip manufacturing costs.
AMD is the leader in Chiplet server chips, with its first generation of Chiplet-based AMDEPYC processors packing eight “Zen” CPU cores, two DDR4 memory channels and 32 PCIe lanes. In 2022 AMD officially releases its fourth generation EPYC processors with up to 96 5nm Zen4 cores and using the next generation Chiplet process, combining 5nm and 6nm processes to reduce costs.
Intel’s 14th generation Core Meteor Lake, which uses the intel 4 process for the first time and introduces the Chiplet small chip design for the first time, is expected to be launched in the second half of 2023, with a performance-to-power ratio target of at least 1.5x the level of 13th generation Raptor Lake.
Part.3
Conclusion
Recently, Intel’s global senior vice president and chairman of China, Wang Rui, said in an interview with the surging news reporter, will study the layout of the class ChatGPT arithmetic model with Chinese customers. “We have very deep cooperation with Baidu and Ali, and the next step of how arithmetic can help us build new models are worth looking forward to.” Under the wave of global digitisation and intelligence, smartphones, autonomous driving, data centres, image recognition and other applications are driving the rapid growth of the AI chip market, and more companies will focus on AI chip production in the future.