半导体行业观察·2026-01-24 02:39Core Viewpoint - The article discusses the complementary relationship between High Bandwidth Memory (HBM) and High Bandwidth Flash (HBF) in addressing the increasing demands of AI workloads on memory systems, highlighting the advantages and limitations of each technology [3][4][5]. Group 1: HBF and HBM Overview - HBF utilizes multi-layer 3D NAND chip stacking technology, which complements HBM for GPU applications [1]. - AI workloads are putting unprecedented pressure on memory systems, necessitating a reevaluation of data delivery to accelerators [3]. - HBM serves as a fast cache for GPUs, enabling efficient reading and processing of key-value (KV) data, but it is expensive, fast, and has limited capacity [3]. Group 2: HBF Characteristics - HBF allows GPUs to access larger datasets while limiting write cycles to approximately 100,000 per module, requiring software to prioritize read operations [4][5]. - HBF's capacity is about ten times that of HBM, but its speed is slower than DRAM [5]. - HBF is expected to debut alongside HBM6, with multiple HBM stacks interconnected to enhance bandwidth and capacity [4]. Group 3: Future Developments - Future iterations like HBM7 may operate as a "memory factory," processing data directly from HBF without traditional storage networks [6]. - A single HBF unit can reach a capacity of 512GB and a bandwidth of 1.638TBps, significantly surpassing standard SSD NVMe PCIe 4.0 speeds [6]. - Companies like Samsung and SanDisk plan to integrate HBF into AI products from Nvidia, AMD, and Google within the next 24 months [6]. Group 4: Market Predictions - The adoption of HBF is expected to accelerate with the HBM6 era, with Kioxia developing a 5TB HBF module prototype using PCIe Gen 6 x8 interface and a transmission rate of 64Gbps [7]. - By 2038, the market size for HBF could potentially exceed that of HBM, according to predictions from industry experts [7].