A new patent from Nvidia shows the company is working on stacking one GPU chip on top of another for more graphics computing power.
Twitter user @Underfox spotted the patent, which was filed in January 2020 and officially granted today on Tuesday. It involves a component called a through silicon via (TSV), which can vertically interconnect one chip die to another. Nvidia is proposing placing the chip dies “face-to-face” against each other, and then linking them together with a number of TSVs.
The technology’s goal is to improve the electrical flow between the two chip dies. “By extending the TSV, resistance may be reduced, allowing for enhanced power delivery to the second semiconductor die,” the patent says. “Resistance may be further reduced by allowing for the TSV to connect to a thicker metallization layer than would otherwise be possible.”
Nvidia has already been using chip-stacking technology to vertically integrate memory on top of GPU chips since 2014 with the Pascal architecture. However, the new patent specifically mentions creating a semiconductor package “wherein the processing unit in the first semiconductor die is a first graphics processing unit (GPU) and the processing unit in the second semiconductor die is a second GPU.”
The obvious benefit to GPU stacking is it could double the transistor density, and thus increase a PC graphics card’s power while maintaining the same overall footprint. However, 3D chips are hard to pull off. A major challenge has been stacking the silicon without introducing defects. The other issue is the risk of overheating with the chip dies stacked so closely together.
Nvidia’s patent focuses on how the same chip stacking requires “hundreds or thousands” of interconnects between the chip dies to efficiently transmit data. “To achieve high speed operation, connection resistance should be low with short path lengths to minimize inductive and capacitive effects,” the company added. “Further, as semiconductor devices scale and more transistors are integrated into a device, the power and current levels required to operate the transistors may also increase—thereby making power delivery more challenging.”
To improve the electrical flow, Nvidia is proposing placing the chip dies in a face-to-face arrangement, instead of a face-to-back configuration.
“A face-to-face arrangement may allow for significantly higher density of interconnects between the dies by forming interconnects from bump pad metal,” the company added. The patent then goes on to say Nvidia could test out each chip die before placing them in a face-to-face configuration, which could help improve manufacturing yields.
Time will tell if Nvidia might actually use GPU chip stacking in an actual product. But back in 2018, major chip foundry TSMC did unveil its own chip wafer-on-wafer technology.