是说芯语·2025-05-07 06:05Core Viewpoint - The rapid advancement of AI large models is driving the transition of AI from a supportive tool to a core productivity force, with computing power chips being crucial for training and inference of these models [2]. Group 1: Computing Power Indicators - Process Technology: Major overseas companies are utilizing advanced process technologies, with Nvidia's latest Blackwell series using TSMC's 4NP (4nm) technology, while AMD and Intel are at 5nm. Domestic manufacturers are transitioning from TSMC's 7nm to SMIC's 7nm [3][4]. - Transistor Count and Density: Nvidia's B200 chip, using Chiplet technology, has a transistor density of 130 million/mm², while Google's TPU Ironwood (TPU v7p) boasts a density of 308 million/mm², significantly higher than competitors [6][7]. - Performance Metrics: Nvidia's GB200 achieves FP16 computing power of 5000 TFLOPS, while Google's TPU Ironwood reaches 2307 TFLOPS, showcasing a significant performance gap [10][11]. Group 2: Memory Indicators - Memory Bandwidth and Capacity: Most overseas manufacturers are using HBM3e memory, with Nvidia's GB200 achieving a bandwidth of 16TB/s and a capacity of 384GB, significantly surpassing domestic chips that primarily use HBM2e [18][19]. - Arithmetic Intensity: Nvidia's H100 has a high arithmetic intensity close to 600 FLOPS/Byte, indicating efficient memory bandwidth usage, while domestic chips exhibit lower arithmetic intensity due to their lower performance levels [20][21]. Group 3: Interconnect Bandwidth - Interconnect Capabilities: Overseas companies have developed proprietary protocols with interconnect bandwidth generally exceeding 500GB/s, with Nvidia's NVLink5 reaching 1800GB/s. In contrast, domestic chips typically have bandwidth below 400GB/s, with Huawei's 910C achieving 700GB/s [26][27].