GPU package
What is a silicon interposer?
This layer provides fine electrical connections between the compute dies and memory. It helps these separate chips work together in one package.
Explore it in the 3D GPU modelWhat it does
This layer provides fine electrical connections between the compute dies and memory. It helps these separate chips work together in one package.
Why it sits where it does
It sits directly beneath the chips and above the package substrate. The interposer connects closely spaced chip contacts; the substrate connects the assembly toward the larger board.
How it works
Small conductive paths carry electrical signals between chips.
Limits and trade-offs
Fine connections need precise manufacturing and assembly. An interposer is one packaging approach, and different products use different layouts and materials.
In Blackwell Ultra and MI300X
NVIDIA Blackwell Ultra
A molded RDL interposer sits beneath the chips. Embedded local silicon bridges provide dense connections near adjacent die edges.
Local silicon bridges handle dense nearby links. Copper redistribution layers and vertical interconnects carry signals and power through the package.
See it in the NVIDIA Blackwell Ultra modelAMD MI300X
A passive silicon interposer sits below the active I/O dies and HBM, above the substrate.
Fine lateral wiring joins I/O dies and memory. TSVs carry connections vertically through the silicon interposer toward the substrate; lateral die-to-die paths need not traverse those TSVs.
See it in the AMD MI300X modelHow it is made
- Form fine conductive paths on silicon, with vertical connections where the design needs them.
- Align and connect the compute and memory chips to that layer.
- Integrate the assembly with the substrate and test its connections.
How it developed
- Early 2010s
Connecting separate chips in one package
TSMC announced a CoWoS test vehicle combining a processor and memory on one interposer.
TSMC · CoWoS test vehicle announcement ↗ - Recent variants
Different ways to build the connection layer
TSMC added two CoWoS variants. CoWoS-R uses redistribution layers; CoWoS-L adds embedded local silicon interconnects, as shown in the Blackwell schematic.
TSMC · CoWoS packaging options and production history ↗
Where it is used
TSMC CoWoS-S
Production technology
TSMC describes a silicon interposer connecting logic chips and HBM. The MI300X schematic shows a passive silicon interposer beneath its active I/O dies.
TSMC · CoWoS packaging options and production history ↗CoWoS-R and CoWoS-L
Alternative structures
R uses polymer and copper routing. L combines a redistribution layer with local silicon connections. Not every advanced package uses a full silicon interposer.
TSMC · CoWoS packaging options and production history ↗
Key terms
- Interposer
- A connection layer placed between chips and their supporting package.
Sources
- TSMC · CoWoS packaging options and production history ↗
- TSMC · CoWoS test vehicle announcement ↗
- NVIDIA · Inside Blackwell Ultra, August 2025 ↗
- Micron · HBM3E stack and interface ↗
- AMD · MI300 architecture and cache hierarchy ↗
- AMD · Instinct GPU specifications ↗
The 3D models are illustrative teaching schematics, not manufacturer CAD. Counts and layouts are simplified where the sources say so.