To fit more HBM into chips, Nvidia accelerates advancement of glass substrate technology

NVIDIA is working together with supply chain companies to advance the implementation of glass substrate technology, in order to address the increasingly prominent physical limitations of AI chip packaging in terms of size, heat dissipation, and interconnection. If there is a breakthrough in this technology, glass substrates are expected to support larger-scale HBM integration, enhancing packaging density and data transfer capacity for the next generation of GPUs.

According to reports, German equipment manufacturer SCHMID revealed at its recent first-half performance briefing that the company is collaborating with key players in the supply chains of Intel, NVIDIA, and AMD to jointly push forward the development of glass core substrate equipment. SCHMID's Chief Strategy Officer Roland Rettenmayer stated that there are still technical obstacles in the glass through-via (TGV) metallization process, and end-customer certification has yet to be completed.

NVIDIA CEO Jensen Huang has previously engaged in multiple discussions regarding this technological approach. During his meeting with SK Group Chairman Chey Tae-won, they discussed next-generation semiconductor cooperation, including glass substrates, and have also actively pushed TSMC to accelerate the development of related technologies. This indicates that NVIDIA is currently assessing the feasibility of introducing glass substrates into the next generation of GPUs.

Glass substrates target AI chip packaging bottlenecks and may expand HBM integration

As AI chip performance and power consumption continue to increase, the limitations of traditional plastic substrates in heat resistance and warpage control are becoming more apparent. Glass substrates feature higher flatness and strength, a lower coefficient of thermal expansion, less deformation under high temperatures, as well as good dielectric properties and lower signal loss. For these reasons, they are viewed as a potential solution for next-generation high-performance chip packaging.

Another major advantage of glass substrates lies in their ability to scale packaging sizes. Current mainstream AI chips typically use a silicon interposer to integrate the GPU and HBM side by side, but silicon interposers have constraints in terms of size scaling and manufacturing cost. Glass substrates are expected to partially or even fully replace silicon interposers, accommodating more chips and HBM in a single package, thus improving packaging density and shortening data transfer paths.

For AI chips, this means greater HBM integration and higher data throughput. HBM, by vertically stacking multiple layers of DRAM to increase bandwidth, is a key memory component in current AI accelerators. As model sizes and computing power demands continue to rise, the amount of HBM that can be integrated within a package has become an important factor limiting AI chip performance.

TGV metallization remains the key to mass production, commercialization not yet achieved

However, there are still technological and certification hurdles to overcome before glass substrates can achieve large-scale commercialization. SCHMID has explicitly pointed out that the TGV metallization process remains one of the main technical barriers at present. Since glass itself does not conduct electricity, forming stable, reliable, and high-density interconnect vias within the substrate is key to achieving mass production.

At the same time, the end-customer certification process is not yet complete, meaning that the related technology is still in the industrialization promotion stage. SCHMID is currently developing the necessary equipment and solutions for glass substrate production, while also expanding its AI server motherboard and next-generation packaging equipment business. For NVIDIA and its supply chain, whether glass substrates can ultimately enter large-scale GPU production still depends on the further maturation of manufacturing processes, reliability, and customer certification.