NVIDIA Considering CoWoP (Chip-on-Wafer-on-Platform PCB) Package Starting With Rubin GR150 GPUs

NVIDIA has reportedly been considering CoWoP as its next packaging solution, possibly being used by its next-gen Rubin GPUs.

NVIDIA Rubin GPUs Could Utilize CoWoP Packaging Solution Instead of CoWoS

CoWoS or Chip-on-Wafer-on-Substrate is the go-to packaging solution for modern-day HPC and AI chips. This technology made its first debut almost 14 years ago, and is currently used by NVIDIA's and AMD's AI powerhouses. The other crucial thing about CoWoS is that it is a mature packaging technology with a robust supply chain that comprises various partners. So one can't go wrong with CoWoS, but it looks like AI giants such as NVIDIA are now planning to switch gears.

According to a roadmap leak (via DigiTimes), it looks like NVIDIA has started looking into CoWoP for its next-generation GPUs. CoWoP or Chip-on-Wafer-on-Platform PCB removes the package substrate and connects the interposer directly to the motherboard.

Some of the key benefits of CoWoP include:

• Better SI with no substrate losses and an increase in NVLINK reach
• Better PI due to better PDN (VR closer to GPU, lower parasitics)
• Better Thermals with lidless and direct die contact
• Reduced PCB Coeff of Thermal Expansion to address warpage
• Improved Electro migration
• Improved AISC cost (No Package, No Lid)
• Better serves the long-term vision of the Dielet model

As you can tell, the Signal and Power integrity are better with CoWoP, reducing substrate losses, and also bringing the voltage regulation closer to the main GPU die. These interfaces also increase NVLINK IC capabilities. The other benefit is that CoWoP packaging doesn't require a package lid, which means that the thermal solution can make direct contact with the silicon. This also leads to reduced costs as you no the need a package and the lid is eliminated.

Based on the leaked roadmap, it looks like NVIDIA started early testing for CoWoP this month. The early test unit is based on the GB100 GPU and features a Dummy GPU/HBM solution. This has a form factor of 110x110mm, and the goal is to evaluate the process flow selection.

Next month (August 2025), NVIDIA is said to begin testing on a functional GB100 CoWoP, featuring a functional GPU/HBM. This unit will retain the same form factor and is designed to evaluate the manufacturability, structural, and electrical functionality, thermal design, and NVLINK interface throughput. An e6540 board with two GB102 GPUs will be used, but there aren't going to be any external customers to evaluate this.

Next year, NVIDIA plans to start test work with its Rubin chips using the same CoWoP package. The GR100 CoWoP will feature an SXM8 form factor, and the goals include Opportunistic POR and pipe clean for GR150. To put it in simple words, the GR100 CoWoP will be the test bed that would pave the way for a fully production-ready GR150 "Rubin" solution.

The Rubin GR150 CoWoP solution will be ready for production around late 2026, which means that it will be available sometime in 2027. But even with CoWoP, NVIDIA isn't going to do away with CoWoS; instead, they will utilize both technologies.

But it's not roses and sunshine with CoWoP. A shift to a completely new packaging solution means that the initial cost of the technology and setting up an entire supply chain will be a huge undertaking, and along with this, CoWoP will add even more complexity and design changes to existing motherboards, so those would also incur a cost rise and production bottlenecks.

Morgan Stanley has also reportedly said that there are little chances of NVIDIA adopting CoWoP for its next-generation GPUs, and several investor firms seem to be backing that. Though it all depends on the market trends and supply/demand factors. It will be a few years before Rubin comes to market, and only then will we be able to see if CoWoP is used; the good 'ol CoWoS is still rocking on Rubin.

News Source: Jukan

Follow Wccftech on Google to get more of our news coverage in your feeds.