Why AMD is pushing chiplets

Are chiplets the future of high-performance computer chips? In a new article, Next Platform explains why AMD is pursuing a chiplet approach for new products. Basically, the cost to make chips has been increasing over the past couple of chip generations, and there are physical limits on the maximum size of silicon dies. Rather than making big monolithic chips, AMD's idea is to pair several smaller chips.

The site points to calculations from AMD, which show it's much cheaper to manufacture four 213mm² chiplets than a single 777mm² monolithic die. Despite the 9.7 percent larger total die area of the four chiplets, caused by overhead, it's estimated the chiplet design would cost just 59% of the cost of the monolithic die.

“Each chiplet had a die area of 213mm2 in a 14nm process, for a total aggregate die area of 4213mm2 = 852mm2 . This represents a ~10% die area overhead compared to the hypothetical monolithic 32- core chip. Based on AMD-internal yield modeling using historical defect density data for a mature process technology, we estimated that the final cost of the quad-chiplet design is only approximately 0.59 of the monolithic approach despite consuming approximately 10% more total silicon.”