随着高带宽内存（HBM）技术向更高层数堆叠演进，散热和良率问题正成为产业链面临的新挑战。最新消息显示，三星电子和SK海力士计划从2027年第二季度开始，大幅增加面向英伟达下一代AI加速器Rubin Ultra的8层HBM4内存供应，而非此前被广泛看好的12层堆叠方案。
报道称，英伟达已于2026年第三季度与三星和SK海力士展开讨论，希望为未来产品引入更多8层HBM4配置。虽然此前市场普遍认为Rubin系列将主要采用更先进的12层HBM4或HBM4E内存，但供应链现实促使公司重新评估这一计划。
业内分析认为，问题的核心并非性能，而是制造难度。随着DRAM芯片层数不断增加，每增加一层都会提升热阻，使散热变得更加困难，同时也增加晶圆键合出现缺陷的概率，导致整体良率下降。相比之下，8层堆叠虽然在单堆容量上有所妥协，却更容易制造和冷却，因此被视为兼顾成本、产能与稳定性的折中方案。
不过，这并不意味着12层HBM4将被放弃。目前三星和SK海力士仍在量产12层HBM4产品，并且已经获得英伟达认证。未来数年内，12层HBM4预计仍将是高端HBM市场的重要组成部分，只是从2027年中开始，8层版本的市场占比有望显著提升。
根据韩国媒体披露的信息，三星和SK海力士还需要为8层HBM4获得独立认证。目前两家公司正在准备相关零部件、材料以及品质验证流程。由于8层方案技术复杂度低于12层方案，业内人士认为后续认证难度相对较低，不太可能再出现重大延误。
值得注意的是，英伟达不仅在评估8层HBM4，也在研究8层HBM4E方案。最终采用何种配置，将取决于Rubin Ultra及其NVL机柜级系统的设计需求、供货能力以及整体成本结构。
与此同时，Rubin Ultra项目本身也出现了一些调整。此前外界曾预计该产品会采用更激进的大规模芯粒设计，但供应链压力已促使英伟达缩减部分规划。据报道，Rubin Ultra已从原先更复杂的方案调整为双芯片设计，以降低制造风险并保证量产进度。
作为英伟达计划于2027年推出的新一代AI计算平台，Rubin和Rubin Ultra将接替Blackwell系列成为公司的旗舰产品。如今HBM4从12层向8层部分回调，也反映出AI产业正在从单纯追求参数和堆叠高度，转向更加重视良率、散热和供应稳定性的现实阶段。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
As high-bandwidth memory (HBM) technology evolves towards higher-level piles, the problems of heat dispersion and good rates are becoming new challenges for the industrial chain. According to the latest information, the Samsung Electronics and the SK Hercules plan, which began in the second quarter of 2027, significantly increased the supply of HBM4 memory on the 8th floor to the next generation of AI Accelerator Rubin Ultra, instead of the previously widely regarded 12-storey stacking programme.
It was reported that British Weida had initiated discussions with Samsung and SK Hercules in the third quarter of 2026, hoping to introduce an additional eight-floor HBM4 configuration for future products. Although it was previously widely accepted in the market that the Rubin series would mainly use more advanced 12-storey HBM4 or HBM4E memory, supply chain realities prompted companies to reassess the scheme.
In industry analysis suggests that the core of the problem is not performance, but difficulty. With the increasing number of DRAM chips, each additional layer increases the heat resistance, making dissipation more difficult, while increasing the probability that the crystals collage with defects, leading to a reduction in the overall benign rate. By contrast, the eight-storey stacking, while compromising on a single stack capacity, is easier to manufacture and cool, and is therefore seen as a compromise between cost, capacity and stability.
However, this does not mean that the 12th floor of HBM4 will be abandoned. Samsung and SK Hercules are still producing 12-storey HBM4 products and have been certified by the British Government. The 12-floor HBM4 is expected to remain an important part of the high-end HBM market in the coming years, with a significant increase in market share expected from mid-2027.
According to information disclosed in the Korean media, Samsung and SK Hercules also need independent certification for HBM4 on the 8th floor. Two companies are currently preparing spare parts, materials and quality verification processes. As the technical complexity of the eight-tiered programme is below the 12th floor, industry believes that subsequent certification is relatively less difficult and unlikely to be significantly delayed again.
It is worth noting that In Weida is not only evaluating the 8th floor HBM4 but is also studying the 8th floor HBM4E programme. The final configuration will depend on the design requirements, supply capacity and overall cost structure of Rubin Ultra and its NVL cabinet system.
At the same time, some adjustments were made to the Rubin Ultra project itself. While it had previously been anticipated that the product would be designed on a larger scale, the supply chain pressure has led to the scaling down of partial planning in Weida. Rubin Ultra was reported to have been modified from a previously more complex programme to a dual chip design to reduce the risk of manufacture and to ensure production progress.
Rubin and Rubin Ultra will take over from the Blackwell series as the company ' s flagship product as a new generation of AI computing platforms planned to be launched in 2027. The current partial return of HBM4 from the 12th to the 8th floor also reflects the shift of the AI industry from a mere pursuit of parameters and stacking heights to a realistic phase of greater emphasis on good rates, heat dispersion and supply stability.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.