In recent years, chip design has had to push back or bypass barriers imposed by physics. The increase in chip power, which is set to accelerate even further with the explosion in demand for AI, means that manufacturers are coming to grips with the limits of the substrates currently in use. Simply put, the substrate is the part (usually green) on which the die of a chip rests. Currently, it’s made of epoxy-type organic resin. As chip size and power consumption increase, the current substrate starts to sweat (almost literally). In concrete terms, it’s buckling under the weight and density of the new chips. It is reaching its limits in terms of electrical connections, just as chip stacking increases this need. It also suffers from the increased temperatures imposed by the new chips. These have to be controlled through dynamic thermal throttling. When overheating occurs, chip performance must be reduced. This brings us to a point of weakness in current substrates: the chip can only maintain optimum performance for a limited time.
Packaging and glass
The substrate is at the heart of an obscure sub-activity that was little talked about just a few years ago: packaging. Packaging has become essential because it’s what makes it possible to overcome, bypass or hack Moore’s Law. It’s what makes it possible to stack more and more chips or chipsets. But what has glass got to do with it?
- Switching to a glass substrate considerably increases the temperature the printed circuit board can withstand. This means the chip can maintain optimum performance over a longer period.
- The glass substrate can be very thin, yet rigid and very flat. These characteristics enable more precise etching, which can increase the density of on-board transistors.
- Glass transmits energy with fewer losses, enabling ten times more electrical connections than current organic substrates.
Glass substrates: Intel holds the key
The news has gone almost unnoticed, but Intel has signed a deal with Nvidia… A surprising situation, given the energy Pat Gelsinger is expending to sell the relevance of his strategy for a strong return to production capacity. While this deal is less high-profile than the one secured for Microsoft’s chip manufacturing, it is nonetheless indicative of the story we developed above. The deal signed between Nvidia and Intel concerns the use of Intel Foundry Services for the production of H100 GPU packaging. Admittedly, Nvidia will continue to entrust 90% of packaging to TSMC, but this open door is not insignificant. It undoubtedly paves the way for the future.
Intel has been working on substrates for years and has made them a key element of its strategy. Intel is even several steps ahead. After more than a decade of collaboration, notably with the French company Soitec, but also with academics, the American company has set up a research center on the subject. Intel’s strength lies in the fact that, over the last few years, it has turned this giant “laboratory” into an operational production line. This production plant is located in Chandler, Arizona. The pilot production line cost almost $1 billion. According to Intel, the use of glass will begin on very high-end chips and gradually become normalized in the second half of this decade.
