NVIDIA Explores CoWoP (Chip-on-Wafer-on-Platform PCB) Packaging for Rubin GR150 GPUs

NVIDIA Explores CoWoP (Chip-on-Wafer-on-Platform PCB) Packaging for Rubin GR150 GPUs

NVIDIA appears to be exploring the potential of CoWoP as its next advanced packaging solution, which may play a significant role in the upcoming generation of its Rubin GPUs.

NVIDIA Rubin GPUs May Implement CoWoP Technology in Shift from CoWoS

For nearly 14 years, CoWoS (Chip-on-Wafer-on-Substrate) has been a pivotal technology in the realm of high-performance computing (HPC) and artificial intelligence (AI) chips, widely adopted by industry giants like NVIDIA and AMD. This packaging method is not only well-established but also benefits from a strong supply chain that involves multiple partners. Despite its proven track record, NVIDIA seems to be contemplating a shift towards CoWoP (Chip-on-Wafer-on-Platform PCB), indicating a strategic move that could redefine its future product lineup.

Recent leaks in development roadmaps disclosed by DigiTimes suggest that NVIDIA is actively testing CoWoP for its next-generation GPUs. Notably, CoWoP technology eliminates the traditional package substrate, allowing direct interposer connections to the motherboard, which could enhance overall performance.

NVIDIA Rubin System specs: 576 GPUs, 2304 memory chips, 12672 CPU cores, 1300T transistors.
Image Source: NVIDIA

CoWoP offers several advantages, such as:

  • Enhanced Signal Integrity (SI) with minimal substrate losses, extending NVLINK reach
  • Improved Power Integrity (PI) due to a more efficient Power Delivery Network (PDN)
  • Superior thermal management with lidless designs and direct die contact
  • Lower Coefficient of Thermal Expansion (CTE) on PCBs, mitigating warpage issues
  • Increased Electro-migration resilience
  • Reduced manufacturing costs by eliminating packaging and lids
  • Alignment with the long-term vision of the Dielet model

The differences in Signal and Power integrity through CoWoP not only minimize substrate losses but also position voltage regulation closer to the GPU die, thus optimizing performance. Furthermore, the absence of a package lid facilitates direct thermal contact with silicon, contributing to reduced costs and an efficient design.

Per the leaked schedule, preliminary testing for CoWoP commenced this month with NVIDIA evaluating the GB100 GPU along with a Dummy GPU/HBM setup. This testing will focus on the selection of manufacturing process flows with a physical form factor of 110x110mm.

In August 2025, NVIDIA is expected to test a functional version of the GB100 CoWoP, integrating an operational GPU and HBM, while maintaining the original dimensions. This phase will assess manufacturability, structural integrity, electrical performance, thermal design, and NVLINK interface throughput, utilizing an e6540 board equipped with two GB102 GPUs without involving external clients.

Timeline and process details of CoWoP assembly with goals and advancements in chip technology.
Image Source: DigiTimes

Looking ahead to next year, NVIDIA plans to advance testing with its Rubin chips utilizing the CoWoP package. The GR100 CoWoP will feature an SXM8 form factor, aiming for opportunistic production readouts and preparatory measures for the GR150 model. Essentially, the GR100 serves as a testbed, potentially leading to a full production-ready GR150 “Rubin”solution.

Projected to enter production in late 2026 and be available by 2027, the Rubin GR150 CoWoP solution promises to enhance NVIDIA’s offerings. Nevertheless, it is important to note that NVIDIA isn’t completely abandoning CoWoS; instead, both technologies will coexist to meet diverse industry needs.

However, the transition to CoWoP is not without challenges. The integration of a new packaging solution involves significant upfront costs and the establishment of a new supply chain. Moreover, the complexity and design modifications required for existing motherboard frameworks could lead to increased costs and potential production slowdowns.

Additionally, Morgan Stanley suggests that the prospects of NVIDIA adopting CoWoP for its upcoming GPUs remain uncertain, with various investor firms echoing similar reservations. The outcome hinges heavily on market trends and supply-demand dynamics. It could take several years for the Rubin line to reach the market, with more clarity on the actual application of CoWoP only becoming apparent once released, while CoWoS continues to be a reliable option for NVIDIA.

News Source: Jukan

Source & Images

Leave a Reply

Your email address will not be published. Required fields are marked *