Advanced Micro Devices dismantled the traditional monolithic processor design by utilizing a microscopic bridge called Infinity Fabric to stitch multiple smaller chiplets…

Why smaller dies can make better processors

A processor’s die is the piece of silicon that contains its circuitry. As a die grows, it covers more wafer area—and has a greater chance of including a flaw that makes the whole piece unusable. That makes large, high-core-count processors expensive to manufacture: one defect can spoil a substantial amount of silicon.

Chiplet design changes the size of the bets. Instead of putting every major function on one large die, a processor can divide the work among several smaller dies. A flaw is then more likely to affect one small piece rather than an entire giant processor. Smaller dies can also be reused in different product configurations, helping manufacturers build a family of processors without designing a new monolithic die for each one.

The connection is part of the design

Those separate pieces still need to behave like one processor. AMD’s Infinity Fabric is the interconnect technology that moves data among parts of its processors; it is not simply a tiny bridge glued between chiplets. The dies are mounted and connected in a package, and the interconnect carries traffic such as requests for data and communication between cores.

That link has to be fast and carefully coordinated. If moving information between dies takes too long, the modular design can lose performance. Engineers therefore have to balance the benefits of smaller manufacturing units against the power, latency, and complexity of communication across the package.

More cores without one enormous die

For server processors, chiplets offer a practical route to adding many cores while keeping each individual die manageable to manufacture. The payoff is not that defects disappear, or that every chiplet automatically runs faster. Rather, more of the silicon made on a wafer can end up in usable products, while manufacturers can scale processor designs by combining dies.

The consequential shift is architectural: a high-end processor no longer has to be one flawless slab of silicon. It can be a coordinated system of smaller pieces, with the package and its data links doing some of the work once confined to a single die.