According to the Korea Herald, the South Korean government plans to collaborate with AI chip and cloud computing companies to form a team to develop high-performance, low-power Neural Processing Unit (NPU) inference chips. The aim of this initiative is to compete with NVIDIA while avoiding the GPU market dominated by NVIDIA.
On June 26, South Korea’s Minister of Science and ICT, Lee Jong-ho, chaired a dialogue on the AI semiconductor strategy, announcing the launch of the first phase of the “K-Cloud Computing” project. Notable chip leaders such as Samsung Electronics and SK Hynix, along with several AI chip startups and cloud computing companies, participated in the meeting.
It is reported that the first phase of the project will invest 100 billion won (approximately 55 million RMB), with the goal of completing the validation of the neural network processor by 2025; the second phase aims to develop low-power Processing In Memory (PIM) chips by 2028; and the third phase targets the development of ultra-low-power in-memory processing chips by 2030.
ASICs and GPUs May Compete as Alternatives
The NPU chip, short for Neural Processing Unit, is a type of AI chip. There are various types of AI chips, such as GPUs, FPGAs (Field Programmable Gate Arrays), and ASICs (Application-Specific Integrated Circuits). NPUs are a type of ASIC chip, as they are custom-designed for specific application scenarios (such as neural networks and deep learning).
NPUs are specifically used for processing artificial intelligence tasks such as neural networks, deep learning, and machine learning, and they have wide applications in AI and deep learning fields, including autonomous vehicles, smartphones, smart home devices, voice recognition, and natural language processing.
It is worth mentioning that the recent emergence of chips like TPUs, NPUs, VPUs, and BPUs are all classified as ASICs. Unlike the flexibility of GPUs and FPGAs, ASICs are customized and cannot be modified once manufactured, resulting in high development costs and long cycles. However, ASICs outperform the former two in terms of performance and power consumption; for example, TPUs can achieve a speedup of 15-30 times compared to contemporaneous GPUs, with energy efficiency improvements of 30-80 times.
Tianfeng Securities believes that as the drawbacks of high power consumption in GPUs become apparent, customized high-performance AI chips (ASICs) may find a market, and in the future, GPUs and ASICs may engage in competitive substitution.
Domestically, companies such as Huawei and Cambricon are laying out NPU chips. Huawei’s NPU products are known as the Ascend series, including Ascend 310 and Ascend 910, primarily used in AI computing; Cambricon’s NPU products mainly include Cambricon 1A, Cambricon 1H, and Cambricon 1M.
According to Zhongtai Securities, the global ASIC market has not yet formed a clear leading manufacturer, and domestic manufacturers are rapidly developing. Domestic products mainly use 7nm process technology, similar to foreign ASIC manufacturers; in terms of computing power, HiSilicon’s Ascend 910 surpasses Google’s latest TPUv4 in BF16 floating-point computing power, and products from Suzhou Technology and Cambricon also match Google’s overall performance. In the future, leading domestic companies are expected to maintain a technological advantage in the ASIC field and break the monopoly of foreign manufacturers in AI chips.