GTC大会前瞻：LPU、CPO和存储

Investment Rating - The industry investment rating is "Positive" [2] Core Insights - The electronic sub-sectors have shown differentiation, with the semiconductor sector increasing by 7.07% year-to-date, while consumer electronics have decreased by 5.22% [3][10] - The upcoming GTC 2026 conference is expected to be a pivotal moment for the "big inference" era, with inference model traffic surpassing 50% since 2025 [24][26] - The report highlights three key technological advancements: LPU (Language Processing Unit), CPO (Co-Packaged Optics), and storage solutions driven by AI computing needs [24][32] Summary by Sections LPU - LPU is a specialized computing architecture optimized for large model inference, emphasizing low latency and high throughput compared to traditional GPUs. NVIDIA's acquisition of Groq is seen as a strategic move to enhance inference capabilities [26][24] - The anticipated introduction of the new Feynman architecture at GTC may incorporate LPU to further optimize AI inference performance, potentially leading to upgrades in server motherboards and PCB designs [26][31] CPO - CPO technology integrates optical modules with switching chips to reduce power consumption and signal loss in high-speed interconnects. The transition to CPO is expected to occur in phases, initially in data center scale-out networks, and later in scale-up applications as technology matures [29][24] - NVIDIA's ongoing investments in optical interconnects suggest a forthcoming launch of scale-up CPO solutions at GTC [29][30] Storage - The demand for high-bandwidth storage is increasing due to AI computing needs, with the next-generation HBM4 expected to enhance bandwidth, capacity, and energy efficiency. The HBM market remains concentrated among a few leading manufacturers, and current production capacity is tight, leading to longer order lead times and rising prices [32][33] - NVIDIA is also optimizing storage architectures for AI inference, which is expected to maintain high levels of industry prosperity [32]