
Multi-Die Package
A multi-die package is a chip housing that connects several individual semiconductor chips — so-called dies — so tightly together that they appear externally as a single chip. This design is today the foundation of the most powerful processors for AI, smartphones, and data centers.
A chip is cut from a thin silicon wafer. The individual piece is called a die (in German sometimes referred to as a “chip core”). For a long time, exactly one such die sat inside a package — the small plastic or ceramic shell that is soldered onto circuit boards. A multi-die package changes that: here, several dies are packed together into a single housing and tightly interconnected. From the outside, the whole thing behaves like a single chip, even though several separate pieces of semiconductor are working inside. In this context, the individual dies are also called chiplets.
Why classic single chips are hitting their limits
Manufacturing a chip means printing circuits onto silicon — with structures that today measure only a few nanometers (millionths of a millimeter) in size. The larger the area to be printed, the more frequently errors occur. A tiny foreign particle on the silicon wafer is enough to render an entire die unusable. Small dies have a much higher yield: more of them are flawless after production and can be sold.
There is also an economic argument. Different parts of a chip — such as the compute core and the memory controller — don’t necessarily need the same expensive manufacturing technology. With a multi-die package, each chiplet can be produced wherever it is cheapest or technically most sensible. This saves AMD costs and development time: a processor's memory controller can be made on an older, cheaper manufacturing line, while the compute cores run on the most advanced process available.
How the dies communicate with each other internally
The central problem is the connection. Two dies sitting side by side on a carrier (the so-called interposer) need to exchange data with each other — and do so as fast as if they were a single chip. For this, thousands of tiny connections are laid, which are shorter and denser than normal solder connections on a circuit board. A common technique is called Through-Silicon Via (TSV): vertical channels are drilled through the silicon, through which electrical connections run. This also allows dies to be stacked on top of each other, not just placed side by side.
Another approach is the interposer, a kind of intermediate layer made of silicon or organic material, on which all the dies are mounted and wired together. Nvidia's H100 chip for AI applications, for example, uses such a design to connect the compute unit and memory with extremely high bandwidth. The data rate between the components is many times higher than with a normal circuit-board connection — which is crucial, because AI models move enormous amounts of data back and forth between memory and the compute core.
A common misconception: multi-die package and multicore are not the same thing. Multicore means that a single die contains multiple compute cores. Multi-die package means that several independent pieces of silicon are assembled together. Both can be combined — and in modern high-performance chips, they usually are.
Multi-die packages in products and news
AMD's Ryzen and EPYC processors have consistently relied on chiplets since 2019. The EPYC server processor “Genoa” consists of twelve chiplets that together offer up to 96 compute cores — no single die could be produced at that size with acceptable yield. Apple's M chips for Mac devices follow a similar idea: the M2 Ultra connects two M2 Max dies via Apple’s own interconnect technology “UltraFusion” so tightly that the operating system and applications address them as a single chip.
In the AI world, multi-die packages have now become standard for high-performance accelerators. Nvidia’s Blackwell architecture (2024) connects two large compute dies on a single carrier, achieving a level of computing performance that simply could not be manufactured with a single die. When news reports talk about new AI chips, there is almost always a multi-die package behind it — even if that is rarely mentioned explicitly.
The technology is also making its way into smartphones: modern mobile memory of the HBM (High Bandwidth Memory) type is built as a stacked die package directly next to or beneath the processor core. The connection density allows data rates that would be unattainable with classic wiring on a circuit board. So this design is no niche topic — it is the foundation on which today’s fastest chips are built.