EB SECURITIES has released a research report indicating that NVIDIA (NVDA.US) plans to incorporate an LPU solution into its new chip architecture, marking a paradigm shift in the AI computing market from "general-purpose computing" towards "specialized inference." The implementation of NVIDIA's LPU solution is expected to significantly increase the required surface area for PCBs and further elevate the complexity of PCB substrate processing. This development is projected to generate substantial incremental demand for PCB micro-drills and PCB processing equipment. With global AI computing demands continuing to grow at a high rate and the need for low latency in AI inference intensifying, the heterogeneous architecture combining GPU and LPU is anticipated to accelerate its adoption. This trend is likely to extend high industry prosperity to the PCB equipment sector, potentially leading to a scenario of supply shortages and price increases for PCB drill bits. The main viewpoints from EB SECURITIES are as follows.

The LPU is characterized by low latency and high bandwidth, forming a complementary relationship with GPUs within AI workflows. The LPU (Language Processing Unit) is a specialized processor designed for AI inference, particularly for low-latency, real-time interactions such as conversations. Its core technology utilizes "compiler-driven" static scheduling to achieve deterministic execution and relies on high-speed on-chip SRAM (with bandwidth reaching up to 80TB/s) to eliminate memory bottlenecks. This reduces initial latency to approximately 100 milliseconds, making it about 10 times faster than the H100 GPU for inferencing mainstream large language models (using Llama2-70B as an example), while also improving comprehensive energy efficiency by about 10 times. In contrast, GPUs (like the H100) are general-purpose, high-throughput architectures that depend on large-capacity HBM memory. They excel at large-scale parallel computations and are the primary choice for large model training and high-throughput tasks. However, in scenarios involving single sequences and real-time generation, they can be constrained by memory bandwidth and runtime scheduling, making it difficult to overcome low-latency bottlenecks. LPUs and GPUs are thus establishing a complementary role in AI workflows: GPUs, with their powerful parallel processing capabilities and large memory, remain central to model training and handling extensive context (the Prefill stage). Meanwhile, LPUs show significant advantages in token-by-token text generation that requires immediate responses (the Decode stage), making them suitable for high-concurrency online inference services. NVIDIA's plan to introduce an LPU solution in its new chip architecture signifies a paradigm shift in the AI computing market from "general-purpose computing" towards "specialized inference."

The incremental impact of LPUs on PCB equipment and drill bits is primarily reflected in two areas: increased PCB value and advancements in packaging. Firstly, the required PCB surface area is expected to increase, accompanied by upgrades in PCB material specifications. Due to the capacity limitations of the 230MB SRAM in a single LPU, running large-scale models may require hundreds of LPUs connected in series. Therefore, if LPUs are adopted on a large scale, the necessary PCB substrate area could increase severalfold compared to architectures using only GPUs. Furthermore, to ensure signal transmission efficiency, LPUs will demand higher-grade PCB materials. It is anticipated that enhanced solutions, such as 52-layer M9-grade copper-clad laminates combined with Q-cloth, will be used, which would consume a substantial number of drill bits. Consequently, the implementation of NVIDIA's LPU solution is projected to significantly boost the required PCB area and further increase the difficulty of PCB substrate processing, leading to considerable incremental demand for PCB micro-drills and PCB processing equipment. Secondly, techniques like PD Disaggregation and 3D stacking may raise the requirements for advanced packaging. At the 2025 GTC conference, NVIDIA introduced PD Disaggregation technology, which splits LLM inference into two phases: the computationally intensive Prefill and the memory-intensive Decode. To address challenges such as insufficient server layout space and increased wiring density from large-scale LPU cluster deployments, NVIDIA is likely to utilize PD Disaggregation to enable the complementary coexistence of GPUs and LPUs, thereby reducing the deployment scale of individual LPU clusters. Simultaneously, the potential use of 3D stacking technology to directly stack LPU units from Groq onto the main GPU chip could help mitigate SRAM capacity shortcomings while preserving low-latency advantages. This would achieve a deep physical-level integration of general-purpose computing and specialized inference. The heterogeneous GPU+LPU architecture demands higher packaging technology and precision, which is expected to further increase the need for high-precision assembly equipment during the PCB electronic assembly process.

Regarding investment targets, it is advisable to focus on equipment manufacturers involved in core PCB manufacturing processes. For high-precision drilling and exposure processes, companies such as Han's CNC (301200.SZ), Inno Laser (301021.SZ), and DR Laser (300776.SZ) are worth noting. For high-precision PCB assembly equipment, attention can be paid to GKG Precision Technology (301338.SZ) and Jintuo Technology (300400.SZ). Concerning high-end PCB drill bits, potential focuses include Diat (301377.SZ), Wald (688028.SH), and SF Diamond (300179.SZ). For advanced plating processes, Dongwei Technology (688700.SH) is a company to monitor. Investors should be aware of risks including intensified industry competition, rapid technological iteration, industrial relocation, and terminal demand falling short of expectations.