THE LIMIT

One exposure, one field

🔲

A lithography scanner prints a chip one field at a time. The largest field is about 26 × 33 mm, or 858 mm². A single die can't be bigger than that, which is called the reticle limit. Top AI dies such as the H100 (about 814 mm²) are already right at it.

AI chips hit the reticle limit years ago, and models kept growing.

TRY IT

Too big for one die?

Choose how much logic you want, and see whether it fits in one reticle field or must be split into chiplets.

Once you pass about 858 mm², the only way to grow is to add dies.

CHIPLETS

Several dies, one chip

📈

Better yield

Small dies are far more likely to come out defect-free

🔀

Mix and match

Compute on the newest node, I/O and cache on cheaper ones

♻️

Reuse

One die design can be used in several products

AMD's MI300X stacks eight compute dies on four I/O dies; NVIDIA's B200 joins two reticle-sized dies.

THE LINK

Making two dies act as one

🔗

Chiplets only work if they can talk almost as fast as wires on one die. The B200's two dies are joined by a 10 TB/s link so software sees a single GPU. Industry standards such as UCIe (Universal Chiplet Interconnect Express) aim to let chiplets from different companies connect.

Die-to-die links need thousands of very short, dense wires, which only advanced packaging can provide.

GOING UP

3D stacking

Chiplets can also be stacked. Hybrid bonding joins copper pads on two dies directly, with pitches of under 10 µm, far denser than solder bumps. AMD's 3D V-Cache stacks extra SRAM on top of CPU dies this way, and the MI300 stacks compute dies on top of I/O dies.

Side by side or on top: either way, the 'chip' is now a package of many dies.

UNIT 10 STUDY COMPLETE

🧩

Ready for the Fab Challenge?

You've seen why one die can't grow past about 858 mm², and how chiplets and fast die-to-die links get around it.