Kools 开发面向 Nvidia Rubin Ultra 级封装的键合技术
半导体封装公司 Kools 开发出一种键合技术,可抑制超大尺寸人工智能(AI)半导体封装的翘曲。该技术采用高刚度方案,面向 Nvidia Rubin Ultra 级封装。原文未披露量产时间、客户导入状态与产能信息。
原文深度解读
本文介绍半导体封装公司 Kools 于 9 月 16 日宣布开发出一种面向超大型 AI 半导体封装的键合技术,该技术将高刚度盖板(high-stiffness lid)与专有的 Thermal Clutch 结合,用于抑制封装翘曲并缩短加热与冷却时间。原文提及的目标应用包括 Nvidia 下一代 AI 加速器 Rubin Ultra 级封装,并给出 180×70mm 超大型封装加热冷却时间从超过 1 分钟降至不到 10 秒、键合后翘曲小于 0.1mm 等关键数字。与电子元器件供应链的关系在于:该技术针对 GPU 与 HBM 多芯片集成的先进封装环节,可能影响超大型 AI 封装的可制造性与 TCB 工艺效率,但原文未披露量产时间、客户导入状态与产能信息。
- Kools 于 9 月 16 日宣布开发出超大型半导体封装键合技术,方案由高刚度盖板与专有 Thermal Clutch 组合构成,目标是抑制超大型 AI 封装在键合过程中的翘曲。
- 工艺顺序发生变化:不是在芯片与中介层键合后的最终封装阶段再贴装用于散热的盖板,而是在键合前先贴装盖板,并对盖板顶部施加压力使封装保持平整,直到焊料固化。
- Thermal Clutch 用于解决高刚度盖板带来的生产率问题:高刚度盖板较厚、热容高,会延长热压键合(TCB)的加热与冷却时间;Thermal Clutch 在加热阶段阻挡热量向冷却段散失,在冷却阶段连通传热路径以快速散热。
- 原文给出的关键数字为:180×70mm 超大型封装的加热与冷却时间从超过 1 分钟降至不到 10 秒,键合后翘曲小于 0.1mm。
- 原文指出,随着封装变大,TCB 正日益取代炉式回流(oven reflow);大封装更易因硅芯片与有机材料之间的热膨胀系数(CTE)差异而变形,导致翘曲和键合缺陷,而 TCB 同时施加热与压力以在键合时保持封装平整,但键合面积增大后加热冷却耗时增加、生产率下降。
- 原文提到 Nvidia 下一代 AI 加速器 Rubin Ultra 是目标应用之一:Nvidia 最初设想在单一封装中集成四个 GPU 裸片和 16 个第七代高带宽存储器(HBM4E),SemiAnalysis 于 6 月报道 Nvidia 因制造顾虑放弃四 GPU 设计并转向两 GPU 结构;Kools CEO Cho Jin-hyun 称其技术可抑制细长超大型封装的翘曲并缩短键合时间,从而改善四 GPU 结构的可制造性。Kools 还表示已获得相关在先专利,知识产权组合共 277 项资产,并计划将该技术扩展至集成 GPU 与 HBM 等多个半导体芯片的下一代先进封装。
供应链影响
- 若该技术按原文所述在超大型封装上实现翘曲小于 0.1mm 并大幅缩短加热冷却时间,可能改善四 GPU 级超大型 AI 封装的可制造性,但实际影响取决于后续量产验证与客户导入情况。
- 由于原文未披露量产时间、客户导入状态与产能信息,该技术对 GPU 与 HBM 先进封装供应链的供给节奏可能仍存在不确定性。
- TCB 在大封装中替代炉式回流的趋势若延续,可能使先进封装环节对热压键合设备、高刚度盖板材料与热管理方案的需求结构发生变化,但具体程度取决于封装尺寸与工艺路线的实际演进。
- Nvidia Rubin Ultra 从四 GPU 设计转向两 GPU 结构的报道,可能反映超大型封装制造难度对产品架构的影响;Kools 技术若被采用,可能为四 GPU 结构路线保留可制造性空间,但这取决于 Nvidia 后续设计决策。
- Kools 计划将技术扩展至集成 GPU 与 HBM 的下一代先进封装,可能使其 IP 组合(277 项资产)在先进封装供应链中的相关性上升,但商业化范围仍取决于客户采用与专利布局的实际执行。
系统解读,仅作为判断线索,不构成备货、出货、涨价或投资建议。
(照片:SemiAnalysis)
半导体封装公司 Kools 开发出一种键合技术,可抑制超大型人工智能(AI)半导体封装中的翘曲。该技术使用高刚度盖板(high-stiffness lid)来控制封装翘曲,并使用 Thermal Clutch 来缩短加热和冷却时间。其设计目的是防止超大型封装(例如 Nvidia 下一代 AI 加速器 Rubin Ultra)中的键合缺陷。
Kools 于 9 月 16 日表示,其通过将高刚度盖板与专有的 Thermal Clutch 相结合,开发出了超大型半导体封装键合技术。该方案不是在芯片与中介层键合后的最终封装阶段再贴装用于散热的盖板,而是在键合前先贴装盖板。对盖板顶部施加压力以使封装保持平整,并维持到焊料固化,从而抑制键合过程中的翘曲。
Thermal Clutch 解决了生产率问题。高刚度盖板较厚且热容高,会增加热压键合(TCB)过程中的加热和冷却时间。Thermal Clutch 在加热期间阻挡热量向冷却段散失。在冷却期间,它连通传热路径,以快速从键合区域散热。
该技术将 180×70mm 超大型封装的加热和冷却时间从超过一分钟缩短到不到 10 秒。键合后的翘曲小于 0.1mm。
随着半导体封装尺寸增大,TCB 在键合工艺中正日益取代炉式回流(oven reflow)。更大的封装更容易因硅芯片与有机材料之间的热膨胀系数(CTE)差异而发生变形,这可能导致翘曲和键合缺陷。炉式回流通过热量熔化焊料,而 TCB 同时施加热量和压力,以在键合期间保持封装平整。
然而,随着键合面积增大,TCB 的生产率会下降,因为加热和冷却需要更多时间。高热输出和翘曲控制是大面积 TCB 工艺的关键挑战。Kools 提出通过用高刚度盖板约束封装,并用其 Thermal Clutch 缩短加热和冷却时间来解决这些问题。
“传统 TCB 需要超过一分钟来加热和冷却 180×70mm 的超大型封装,但 Thermal Clutch 可以将其缩短到不到 10 秒,”Kools 首席执行官 Cho Jin-hyun 表示。“通过在高刚度盖板保持封装形状的同时缩短键合时间,我们可以为超大型四图形处理单元(GPU)封装确保生产率。”
Nvidia 下一代 AI 加速器 Rubin Ultra 等超大型封装属于目标应用之列。Nvidia 最初设想 Rubin Ultra 将四个 GPU 裸片和 16 个第七代高带宽存储器(HBM4E)器件集成到单一封装中。半导体市场研究公司 SemiAnalysis 在 6 月报道称,Nvidia 因制造顾虑放弃了四 GPU 设计,并转向两 GPU 结构。
“Nvidia 最初的 Rubin Ultra 设计将四个 GPU 排成一列置于单一封装内,以便它们可以像单个芯片一样并行运行,”Cho 表示。“应用 Kools 的技术可以在细长的超大型封装中抑制翘曲并缩短键合时间,从而改善四 GPU 结构的可制造性。”
Kools 表示,其已获得该技术的在先专利。其知识产权(IP)组合总计包括 277 项资产。该公司计划将该技术的应用扩展到集成多个半导体芯片(包括 GPU 和 HBM)的下一代先进封装。
(Photo: SemiAnalysis)
Semiconductor packaging company Kools has developed bonding technology that suppresses warpage in ultra-large artificial intelligence (AI) semiconductor packages. The technology uses a high-stiffness lid to control package warpage and a Thermal Clutch to shorten heating and cooling times. It is designed to prevent bonding defects in ultra-large packages such as Nvidia's next-generation AI accelerator Rubin Ultra.
Kools said on Sept. 16 that it developed the ultra-large semiconductor package bonding technology by combining a high-stiffness lid with its proprietary Thermal Clutch. Instead of attaching the lid for heat dissipation at the final packaging stage after bonding chips and interposers, the lid is attached before bonding. Pressure is applied to the top of the lid to keep the package flat and maintained until the solder solidifies, suppressing warpage during bonding.
The Thermal Clutch addresses the productivity issue. High-stiffness lids are thick and have high heat capacity, increasing heating and cooling times during thermocompression bonding (TCB). The Thermal Clutch blocks heat from escaping to the cooling section during heating. During cooling, it connects the heat-transfer path to rapidly dissipate heat from the bonded area.
The technology reduced the heating and cooling time for an ultra-large 180×70mm package from more than one minute to less than 10 seconds. Warpage after bonding is less than 0.1mm.
As semiconductor packages grow larger, TCB is increasingly replacing oven reflow in bonding processes. Larger packages are more susceptible to deformation caused by differences in the coefficient of thermal expansion (CTE) between silicon chips and organic materials, which can result in warpage and bonding defects. Oven reflow melts solder with heat, while TCB applies heat and pressure simultaneously to maintain package flatness during bonding.
TCB, however, becomes less productive as the bonding area increases because more time is required for heating and cooling. High thermal output and warpage control are key challenges in large-area TCB processes. Kools proposes addressing these issues by restraining the package with a high-stiffness lid and shortening heating and cooling times with its Thermal Clutch.
"Conventional TCB requires more than one minute to heat and cool an ultra-large 180×70mm package, but the Thermal Clutch can reduce that to less than 10 seconds," Kools CEO Cho Jin-hyun said. "By maintaining the package shape with a high-stiffness lid while shortening bonding time, we can secure productivity for ultra-large four-graphics-processing-unit (GPU) packages."
Ultra-large packages such as Nvidia's next-generation AI accelerator Rubin Ultra are among the target applications. Nvidia originally envisioned Rubin Ultra with four GPU dies and 16 seventh-generation high-bandwidth memory (HBM4E) devices integrated into a single package. Semiconductor market research firm SemiAnalysis reported in June that Nvidia had scrapped the four-GPU design over manufacturing concerns and switched to a two-GPU structure.
"Nvidia's original Rubin Ultra design placed four GPUs in a row within a single package so they could operate in parallel like a single chip," Cho said. "Applying Kools' technology can suppress warpage and shorten bonding time even in elongated ultra-large packages, improving the manufacturability of a four-GPU structure."
Kools said it has secured prior patents for the technology. Its total intellectual property (IP) portfolio comprises 277 assets. The company plans to expand the technology's application to next-generation advanced packaging that integrates multiple semiconductor chips, including GPUs and HBM.