Crafting Computers with Chiplets

Today’s world is filled with creativity and intelligence, causing products and services to often fuse these two characteristics together, ultimately redefining the world of technology. Though the advancements in technology are impressive, they often have detrimental effects to the environment. However, over the past two years, the Defense Advanced Research Projects Agency (DARPA), an agency of the United States Department of Defense, has been working on a chiplet system to make information systems more eco-friendly.

The main premise of a chiplet is much like a lego system, in which there are smaller modular chips that connect using a die-to-die interconnect scheme.

“If we can truly have an ecosystem of hundreds or thousands of tiny chiplets sitting on the shelf, then we can talk about having days to weeks of state-of-the-art design,” said Andreas Olofsonn, the Program Manager in the Microsystems Technology Office at DARPA.

While DARPA is providing a new and updating way to use technology, the idea of a chiplet is not new. Through a different model, the multi-chip modules are used by integrating dies and connecting them to a larger module. Unfortunately, the multi-chip modules took longer to develop and the traditional scaling was adopted instead.

The chiplet system provides a cheaper and more efficient option than traditional chip scaling. In the traditional scaling, the individual features of the device are shrunk at each particular node. However, this increases the amount of resources needed for the project to function, thus costing more money and requiring more capital. For example, the traditional scaling design increased from $53.1 million for a 28nm planar device to $297.8 million for a 7nm chip.

With this increase in cost, the industry needed to look for different ways to build processing technologies. So, DARPA worked tirelessly to create the chiplet approach. Essentially, 100 IP blocks are divided into smaller chiplets. Then, with a large catalog of chiplets, a designer can mix and match these chiplets to build a system, or piece of technology; they can also be used in different designs. The re-use and mix and matching system decreases the cost, design and effect on the environment of a piece of processing technology. Within the next few years, this design will hopefully be adopted on a larger scale, maybe by the military or aerospace engineers.

While the chiplet model provides a great sense of hope and excitement for the next generation of developments, there are still a plethora of problems and challenges that must be addressed. Currently, there is no standard way of assembling the chiplets. Since there is no generalized key to package the chiplets, the die-to-die interconnect schemes can be expensive. Though this theory is less expensive than the traditional scaling, there is still a large amount of capital required to make this successful. As this is a newer model, investors are hesitant to pour in resources as they are wary of the longevity of the chiplet model. Moreover, as each chiplet model is extremely unique, there is no true test to measure the validity and accuracy of each chiplet. A large part of the problem is that each chiplet approach requires multiple supply chains. However, there is no single formula to ensure that all chiplets will function on their own as well as with other chiplet models. The lack of cohesion and unity creates doubts among leaders of different models. If these individual leaders run into a problem with a particular chiplet model, there is no way to look at other models to see where the error lies.

DARPA recognizes these complications and is ready to work to remedy and improve upon the chiplet model. Oloffson is determined to “reinvent the ecosystem and invent manufacturing technologies.” Since the start of the program in 2017, DARPA is formulating standards for chiplets, devising faster SerDes technologies and developing low-cost bridges for the die-to-die interconnects, the premise of the chiplet model. There is currently a $1.5 billion investment to ensure the success of the CHIPS program. Nevertheless, if the errors are not ironed out in the near future, the chiplet model could very well be another niche programming that never took flight.