NVIDIA's Rubin Ultra to Stick with ABF Substrates, Avoiding CoWoP Transition

Despite the market's fervent discussion around chip-level wafer-on-package (CoWoP) technology, the likelihood of NVIDIA adopting this technology for its next-generation GPU product, Rubin Ultra, is relatively low. On July 30, a report indicated that NVIDIA's Rubin Ultra will continue to use the existing ABF substrate technology rather than transitioning to the CoWoP solution.

Analysts believe that the shift from CoWoS to CoWoP presents significant technical challenges, making it difficult to reduce dependence on ABF substrates in the short term. The report highlights that the complexity of the technology transition and the risks associated with supply chain reorganization make it impractical to widely adopt CoWoP in the near future. However, it does not rule out the possibility that NVIDIA is concurrently developing CoWoP technology as a supplementary measure to address specific issues.

The primary reason for Rubin Ultra's unlikely adoption of CoWoP is the high technical barrier posed by the technology. CoWoP requires the line/space (L/S) of the printed circuit board (PCB) to be reduced to below 10/10 microns, which is comparable to the current standard for ABF substrates. Currently, high-density interconnect (HDI) PCBs have an L/S of 40/50 microns, and even the PCBs used in advanced motherboards only achieve 20/35 microns. Reducing the L/S from 20/35 microns to below 10/10 microns presents a significant technical challenge.

Research indicates that the next-generation GPU will continue to rely on traditional packaging methods, with both Rubin and Rubin Ultra expected to use ABF substrates. The ABF substrate for Rubin Ultra will be larger and have more layers compared to Rubin, which is contrary to the technological path of CoWoP.

In addition to the technical complexities, transitioning from CoWoS to CoWoP would introduce significant risks to yield rates and necessitate a reorganization of the supply chain. Given that the target product is expected to enter mass production within a year, such a technological shift does not make commercial sense. Currently, the yield rate for CoWoS is nearly 100%, and transitioning from this high yield rate to a new technology would introduce unnecessary risks.

The technology transition involves not only changes in manufacturing processes but also the reconfiguration of the entire supply chain ecosystem, which presents high complexity and risk in the short term.

Despite the short-term impracticality of widespread adoption, CoWoP technology holds potential advantages. These include solving substrate warpage issues, increasing NVLink coverage on the PCB without requiring additional substrates between the chip and PCB, achieving higher thermal efficiency without a packaging lid, and eliminating capacity bottlenecks for certain packaging materials. Additionally, it does not rule out the possibility that NVIDIA is concurrently developing CoWoP technology as a supplementary measure to address specific issues such as substrate warpage, material supply constraints, or simplifying the GPU board manufacturing process.