How to Build Giant Computer Chips Beyond Lithography Limits
This patent describes a method for building very large, powerful computer chips by combining smaller 'chiplets' on a single wafer, connecting them seamlessly even when the total chip size is bigger than what a single manufacturing step can create.
Patent Number
US 12431423
Status
Active
Filing Date
June 30, 2023
Grant Date
September 30, 2025
Expiration
June 30, 2043
Claims
23
Assignee
Taiwan Semiconductor Manufacturing Co TSMC
Inventors
Murat Kerem Akarvardar, Hon-Sum Philip Wong
Citations
0 forward · 10 backward
What it covers
This patent details a method for manufacturing a semiconductor device that integrates multiple smaller circuits, called chiplets, onto a single semiconductor wafer. The process involves forming a first circuit of a first chiplet and a second circuit of a second chiplet on the wafer, separated by a 'die-to-die space' filled with a dielectric material. Crucially, the overall 'chip region' can be larger than the maximum area a standard lithography machine can expose in one go, as stated in claim 2. The method then forms a 'conductive connection pattern' directly over this die-to-die space, electrically linking the chiplets so they function as one large circuit (claim 1). For example, a powerful processor could be built by combining separate CPU core chiplets and memory controller chiplets on the same wafer, with high-speed connections formed between them using this technique.
What it doesn't cover
- —Does not cover traditional multi-chip modules where pre-fabricated, separate chips are assembled onto a package substrate, as this method forms chiplets on a single wafer.
- —Does not cover methods where the entire large chip is formed in a single lithography exposure step, as it specifically addresses designs larger than a maximum exposure area (claim 2).
- —Does not cover connecting chiplets using only wire bonding or micro-bumps if those connections are not themselves formed using lithography over the die-to-die space (claim 1, claim 6).
- —Does not cover chiplet designs where the connection patterns are only within the boundaries of each chiplet and do not span the space between them (claim 1 specifies 'disposed over the die-to-die space').
- —Does not cover methods that do not involve using multiple photo masks and exposures to create the different chiplet patterns (claim 1 details first and second exposures with different masks).
The clever bit
The clever bit is the ability to stitch together multiple smaller circuit patterns, each created with its own lithography exposure, onto a single semiconductor wafer, and then form high-bandwidth electrical connections *over the spaces between them* using further lithography steps, all while staying within the limits of existing manufacturing tools.
Why it matters
This patent is significant because it provides a way to overcome a fundamental limitation in semiconductor manufacturing: the 'reticle limit,' which is the maximum size a single lithography exposure can print. By enabling the creation of larger, more complex integrated circuits from smaller chiplets on a single wafer, it allows for the development of more powerful processors and specialized chips without requiring entirely new, larger, and more expensive lithography equipment. This approach is critical for advancing high-performance computing, artificial intelligence, and other demanding applications.
Real-world examples
- 1.High-performance computing (HPC) processors
- 2.Artificial intelligence (AI) accelerators
- 3.Advanced CPUs and GPUs with chiplet designs
- 4.Data center processors
- 5.Specialized application-specific integrated circuits (ASICs)
Generated by PatentBrief · Not legal advice · patentbrief.org
US 12431423 · 2026