SK Hynix has reportedly reached a significant milestone in its pre-production trials by raising the yield rate of its sixth-generation high-bandwidth memory (HBM4) to 70%. This achievement strengthens the South Korean semiconductor company's position in the competitive HBM4 market, as industry observers eagerly await its next moves.
According to South Korea's ET News, citing industry sources, SK Hynix's 12-layer HBM4 has attained a 70% yield in testing—a metric that heralds robust production efficiency once mass manufacturing commences. The company has reportedly imposed exacting standards across each phase of the 12-layer HBM4's testing process. Insiders note that by the close of 2024, the yield surpassed 60%, and recent strides have further bolstered this figure, with improvements unfolding at a notably brisk pace.
The 12-layer HBM4 is said to leverage SK Hynix's fifth-generation 10nm-class DRAM (1b) process node, a technology already proven for its performance and stability in the fifth-generation HBM (HBM3E). This established foundation facilitates a smoother path to consistent yields. Korean media outlets project that the 12-layer HBM4 is poised for a swift transition to mass production. Following the completion of proprietary technology development and evaluation, SK Hynix plans to produce samples for client validation, paving the way for full-scale manufacturing upon approval.
Speculation abounds that this 12-layer HBM4 is tailored for Nvidia's forthcoming AI accelerator, the Rubin series. Initially slated for a 2026 debut, the Rubin timeline may accelerate, with some analysts predicting production as early as the latter half of 2025. According to Korea's News1, SK Hynix intends to initiate large-scale production of its 12-layer HBM4 in the second half of this year, with plans to supply 16-layer variants by mid-2026.
Collaboration with TSMC
The shift to HBM4 introduces heightened complexity, as the base chip linking GPUs and HBM now requires advanced logic processes to accommodate bespoke client specifications. To address this, SK Hynix has opted to collaborate with TSMC, harnessing the Taiwanese foundry's cutting-edge logic frontend capabilities rather than producing the base chips in-house.
In an interview published on February 26 by SK Hynix's news center, Han Kwon-hwan, vice president of HBM Integration Technology, underscored the escalating challenges ahead. He noted that this year's flagship 12-layer HBM3E, compared to its 8-layer predecessor, demands greater technical prowess—a trend that will only intensify with successive HBM generations.
Analysts surmise that SK Hynix's concerted efforts to stabilize the 12-layer HBM4 yield are likely driven by the needs of its marquee client, Nvidia. Should SK Hynix succeed in aligning production with Nvidia's accelerated schedule, it could deliver the 12-layer HBM4 ahead of expectations, reinforcing its stature in the high-stakes realm of advanced memory solutions.
Article edited by Jingyue Hsiao
