Every time you use ChatGPT or generate an image with AI, there is a memory chip working at extreme speed behind the scenes. However, that chip has a memory bottleneck problem, and a Korean research team may have just solved it.
Researchers at POSTECH (Pohang University of Science and Technology) developed a new way to stack more than 10 ultrathin semiconductor chips on top of each other, achieving a memory density roughly four times higher than the best commercial chips available today (via TechXplore).
Why is stacking chips so hard, and what makes this one different?
High-bandwidth memory, or HBM, is the type of memory that powers AI accelerators. It works by stacking multiple chips vertically, much like building a high-rise instead of spreading out across land.
The problem is that as chips get thinner, they become incredibly fragile. At one-fifth the thickness of a human hair, they bend, warp, and crack under pressure. Current manufacturing methods make this worse, often damaging chips before they even make it into a stack.
The POSTECH team solved this by combining two techniques into one process. Transfer printing precisely places each chip where it needs to go, while in-situ bonding forms the metallic connections at the same moment, all under low heat below 180 degrees Celsius and low pressure below 20 kilopascals. The result is a stack of more than 10 chips with almost no misalignment and very little warping.
Why this matters for the future of AI
More memory packed into the same space means AI tools can run faster and handle bigger tasks without needing larger or more expensive hardware. The researchers also see uses beyond AI, including next-generation micro-LED displays and advanced processor designs that need the same kind of ultra-precise stacking this method delivers.
Getting this into commercial production is the next step, but if it gets there, the memory ceiling that has been quietly holding AI back could finally start to lift.