英伟达GB10，深度解读

Core Insights - The GB10 product, a collaboration between NVIDIA and MediaTek, integrates NVIDIA's Blackwell architecture into its GPU, featuring 48 Blackwell SM units and a powerful CPU subsystem with 10 Cortex X925 and 10 Cortex A725 cores, indicating strong performance capabilities [1][3]. Memory Subsystem Analysis - The GB10's CPU cores are organized into two clusters, each containing five A725 and five X925 cores, with A725 cores operating at 2.8 GHz and X925 cores reaching up to 4 GHz [3]. - The A725 cores have a 512 KB L2 cache with a latency of approximately 9 clock cycles, while the X925 cores feature a 2 MB L2 cache with a latency of 12 clock cycles, showcasing a balanced cache configuration for performance [5][7]. - The system-level cache (SLC) of 16 MB serves as an L4 cache for the CPU, enabling efficient data sharing between CPU and GPU without needing to access DRAM, which is a significant feature of the GB10 architecture [8]. Performance Metrics - The GB10's A725 core can read approximately 55 GB/s from the L3 cache, while the X925 core can maintain nearly 90 GB/s, indicating superior bandwidth capabilities [14]. - In multi-threaded workloads, the GB10's memory hierarchy shows that the shared components face increased pressure as more cores are activated, with the total L2 cache of 15 MB and L3 cache of 8 or 16 MB [16]. - The GB10's latency performance is highlighted by its lower baseline latency compared to competitors, achieving lower latency under bandwidth load conditions [28]. Architectural Considerations - The design of the GB10 suggests a focus on performance optimization in cluster 1, while cluster 0 emphasizes density, with a trade-off in L3 cache capacity to save area [20]. - The integration of CPU and memory controller on the same chip may contribute to reduced latency, enhancing overall performance [12]. - The architecture's ability to manage bandwidth demands effectively is crucial, especially when considering the impact of integrated graphics on CPU performance [31][33].