"Computing power infrastructure" focuses on hardware systems, focusing on broad computing power on the one hand, and supporting infrastructure on the other. Among them, the generalized computing power mainly covers the three aspects of computing power chip (CPU + XPU), packaging and testing, and storage, and the supporting infrastructure mainly covers AI servers, supporting cabinets, communication equipment, etc.
The complexity of deep learning models and recommendatory system models has been further improved, which puts forward higher requirements for chip computing power, and AI chip computing power has entered a stage of vigorous development at this stage.
AI operation refers to the neural network algorithm represented by "deep learning", which requires hardware with efficient linear algebra computing ability, and the computing task has the characteristics of simple unit computing task and low difficulty of logic control, but large amount of parallel operation and many parameters. High requirements are put forward for multi-core parallel computing, on-chip storage, bandwidth, and low-latency memory access.
From the perspective of technical architecture, computing power chips are mainly divided into three categories: GPU, FPGA, and ASIC. Among them, GPUs are more mature general-purpose AI chips, while FPGAs and ASICs are semi-custom and fully customized chips for AI demand features. #Artificial Intelligence ##Computing Power ##芯片#
Source: China Academy of Information and Communications Technology
Pay attention to Leqing and gain insight into the industrial pattern!
GPU
GPU was the first concept proposed by NVIDIA when it released the NVIDIA GeForce 256 graphics processing chip in August 1999. Previously, the display chip that processed the image output in a computer was rarely regarded as a separate computing unit.
GPU chips are highly versatile, and a large number of arithmetic logic units (ALUs) originally designed for graphics computing can provide good acceleration effects for deep learning.
GPUs have the most computing power, but they have high cost and high power consumption.
The global GPU market is a three-legged oligopolistic competition pattern, and NVIDIA is the largest in the field of independent display.
Discrete GPU, that is, discrete graphics card, needs to be plugged into the corresponding interface of the motherboard, has a separate video memory, does not occupy system memory, and can provide better display effect and running performance.
Integrated GPU, or integrated graphics card, is to integrate the graphics core on the motherboard as a separate chip, and dynamically share part of the system memory as video memory, which can provide simple graphics processing capabilities and smoother coding applications. The world's well-known suppliers mainly include Intel and AMD.
In the discrete graphics card market, it has long been AMD and NVIDIA two people turn, in 2022 Intel officially entered the graphics card market, the current independent GPU market is mainly occupied by NVIDIA, AMD and Intel three companies, 2022 Q4 global independent GPU market share of 85%, 9% and 6%, respectively, of which NVIDIA in the PC side independent GPU market share advantage is obvious.
IDC expects China's GPU market to reach $11.1 billion in 2023 and $34.557 billion in 2027.
FPGA
Similar to GPUs, FPGAs are also accelerated with CPUs. Unlike CPUs and GPUs, FPGAs can flexibly compile the chip hardware layer and consume much less power than the first two.
The configuration flexibility provided by FPGA programmability allows it to adapt to the market more quickly and has obvious practicality.
FPGAs were invented in 1985 by Ross Freeman, one of the founders of Xilinx, and are further developed on the basis of existing programmable devices such as PAL, GAL, CPLD, etc. FPGAs are mainly composed of programmable I/O units, programmable logic units, programmable wiring resources, etc.
As a semi-custom circuit in the field of application-specific integrated circuits (ASICs), FPGAs can be reprogrammed multiple times to achieve specific functions according to the user's needs through the accompanying EDA software.
Where FPGAs stand in the integrated circuit industry:
The PGA chip consists of three parts: logic unit (LC), input and output unit (IO), and switching line array (SB).
In terms of market structure, overseas manufacturers dominate the global FPGA market, Xilinx and Intel form a duopoly, and domestic enterprises have a first-mover advantage including Fudan Microelectronics (leading high-reliability FPGA technology, the first to launch billion-gate FPGAs and PSoC chips, and the application fields are constantly enriched) and Unigroup Guowei (a leader in the domestic special integrated circuit industry, with products covering more than 500 varieties, and FPGAs in special fields continue to be updated), Anlu Technology (domestic civil FPGA leader).
Fudan Microelectronics has 10 million gate products in 65nm process and 100 million gate level products in 28nm process, and currently mainly focuses on FPGA products with 28nm process.
Unigroup launched a 2x nm low-power FPGA series in 2022, and a new generation of 1x nm higher performance FPGA series products are also progressing smoothly, further improving the product category.
The FPGA chip products of Anlu Technology have formed a product matrix consisting of the PHOENIX high-performance product series, the EAGLE high-efficiency product series, and the ELF low-power product family.
According to MRFR forecasts, the global FPGA market is expected to reach $12.5 billion in 2025.
ASIC
ASIC chips are mainly used in deep learning acceleration, and on the inference side, they have obvious advantages in efficiency and speed compared with other AI chips.
ASIC is small in size, low in power consumption, suitable for mass production, but long development time, and non-editable, high upfront investment cost, bringing certain technical risks.
The most prominent ASICs are Google's TPU (Tensor Processing Chip) released in 2015 and Intel's Gaudi 2 released in 2022.
According to KBVResearch report, the global ASIC chip market size is expected to reach $24.7 billion from 2019 to 2025, growing at a CAGR of 8.2% during the forecast period.
The main AISC enterprises in the mainland include Cambrian, Montage Technology, Black Sesame, Horizon, Huawei HiSilicon, Alibaba, etc.
Some domestic AISC technologies have reached international leadership, such as in terms of BF16 floating point computing power, Huawei's HiSilicon Ascend 910 has surpassed Google's latest generation product TPUv4.
In addition, in edge computing, AI chips will use the data collected by sensors such as microphone arrays and cameras to reason according to the built model and output corresponding results. Due to the many application scenarios of edge computing, the performance requirements for hardware such as computing power and energy consumption are also different, which has spawned a broader demand for AI chips.
The United States has restricted the sale of the most advanced and widely used AI training GPUs - NVIDIA A100 and H100 to China, and there is a big gap between domestic computing power chips and NVIDIA's latest products, but the reasoning operation with lower requirements for information granularity can be partially replaced.
In terms of training chips, leading enterprises focus on the research and development of related chips to break the long-term monopoly of foreign enterprises. At present, the mainland's AI chip industry is still in its infancy, and under the background of the outbreak of a new round of artificial intelligence, the market space of domestic substitution acceleration industry is sufficient.
Pay attention to Leqing and gain insight into the industrial pattern!