都在抢HBM 4

Core Viewpoint - Nvidia is pushing its memory suppliers to exceed the JEDEC official HBM4 benchmark, aiming for a stack speed of 10Gb/s to enhance GPU bandwidth beyond AMD's upcoming MI450 Helios system [2][3]. Group 1: Nvidia's Strategy and HBM4 Development - Nvidia has requested a stack speed of 10Gb/s for its 2026 Vera Rubin platform, which would increase the total transmission rate per stack to 2.56 TB/s, allowing a single GPU to achieve 15 TB/s of raw bandwidth [2]. - The Rubin CPX configuration is designed for demanding inference workloads, reportedly achieving a transmission rate of 1.7 PB per second across the NVL144 rack [2]. - However, achieving 10Gb/s HBM4 is not guaranteed, as higher I/O speeds lead to increased power consumption and stricter timing requirements, which could pressure the underlying chip [3]. Group 2: Supplier Landscape and Competition - SK Hynix remains Nvidia's primary HBM supplier and has completed HBM4 development, preparing for mass production, but has not disclosed specific chip specifications [3]. - Samsung is more aggressive in node migration, transitioning its HBM4 base chip to a 4nm FinFET process to support higher clock speeds and lower power consumption [3]. - Micron has confirmed HBM4 samples with a 2,048-bit interface and bandwidth exceeding 2 TB/s, but has not disclosed support for 10Gb/s [3]. Group 3: AMD's Position and Memory Specifications - AMD's MI450 is still in planning, but it is expected to support up to 432GB of HBM4 memory per GPU, potentially allowing AMD to catch up or surpass Nvidia in raw capacity [4]. - With the CDNA 4 architecture upgrade, AMD aims to leverage the inference advantages of Nvidia's Rubin platform [4]. Group 4: Samsung's HBM3E Certification and Future Plans - Samsung has recently received Nvidia's certification for its fifth-generation 12-layer HBM3E product, marking a significant milestone in the high-bandwidth memory sector [5][6]. - This certification comes after Samsung redesigned the HBM3E DRAM core to address thermal performance issues, which had hindered earlier versions [6]. - Samsung plans to supply HBM4 samples to Nvidia soon, with mass production expected to begin in the first half of 2026 [10].