< Back

Chiplet Summit 2024: Opportunities, Challenges, and the Path Forward for Chiplets

Print Friendly, PDF & Email
“Heterogeneous technology” and “disaggregation” were key terms used in keynotes, tutorials, panels, and individual sessions throughout the Chiplet Summit. Simply put, it is how to take something that was once whole and break it down into many pieces. As you start to “disaggregate” or “cut up” a monolithic chip/die like a complex CPU or GPU you end up with many different pieces. These pieces are now independent and can be used in many ways. What defines how these pieces are used is dependent on the use case and application. These independent pieces are what we call Chiplets.

The ingenuity of Chiplets and their multitude of use cases and designs has also introduced challenges that can turn the benefits and appeal of Chiplets into something counterintuitive and not useful.  Hence, the increased focus around standardization and the effort to create an even playing field to promote adoption.

Chiplets growing popularity

Let’s begin with outlining Chiplet benefits, as every keynote speaker did in their first few slides, in no particular order.

  • Faster time to market
  • Portfolio expansion
  • Higher margins
  • R&D efficiency

Market research firm Yole showed the following data to explain the value of Chiplets and I think it captured the overall message of the conference:

The obvious reason why companies want to use Chiplets is the overwhelming economic value. Originally, companies had to spend millions of dollars on one chip that could still potentially require additional millions of dollars for revisions and still encounter further cost increases due to large size and bad yields. These companies’ chips can now shrink in size, have better yields, and get to market faster with higher margins.  Who wouldn’t want to save money and make money at the same time?

What markets can take advantage of Chiplets?

Today the largest market adopter is the server market. This is due largely to the growth of the datacenter and the strains created by Artificial Intelligence and Machine Learning. AMD, NVIDIA, Intel, and other CPU and GPU products are becoming power hungry, have more compute power requirements, require more memory, are increasing in size, and have yields that are getting worse.  The ability to disaggregate these products and simplify them by making them modular is a solution to these problems. 

In their presentation, Yole separated out specific technologies and markets that are adopting Chiplets, but ultimately, they are all server related as HBM (high bandwidth memory) is now being used as the memory of choice for AI applications, Datacenters, and Generative AI applications.

After server adoption, Yole sees a move in adoption of Chiplets in the PC market, then smart phones and automotive.  Although it should be noted that in automotive there is an added layer of complexity of test and standard requirements that will slow the adoption rate and make adoption more complicated. And smart phones have an overall cost threshold that cannot be crossed.

With Chiplets, what makes them great is also what makes them not so great. Chiplets are great in their modularity but each chiplet module is an independent product with each Chiplet having the ability to be made in different ways, with different characteristics and specifications. With every Chiplet created, each requires their own independent set of quals and productization. This is not a one size fits all solution and there needs to be careful consideration on cost, complexity, and advantages to creating a product using Chiplets.  With each different variation of Chiplet, the process resets itself and a brand-new analysis and qualification process begins.

Creating an open ecosystem

So how do you simplify this and create an ecosystem in which anyone can enter?  At the Chiplet summit everyone was there to promote options. There were two standards committees there to promote their standardization work. The UCIe consortium and OCP (Open Compute Project).  The UCIe consortium is focused on interconnect and OCP is focused on building an “Open Chiplet Economy”. There were packaging companies focused on 2.5D and 3D packaging technologies, as well as how to stack dies using TSV (through silicon via) and companies discussing interposers and CoWos technology. But even with these potential options and offerings, it is too early to know what the right solution might be. For now, the best solution is the one that fits the need of the end user.

Taking this a step further, most attendees had no idea what they were there for or what to ask. Our very own Jeff Twombly, Credo VP of Business Development participated in a panel titled, “Best ways to Optimize Chiplets”. His first question before talking was, “who in the audience has had experience working with Chiplets to date?” Of the 30-40 attendees, only about 6 people raised their hands. When you look at it from a percentage standpoint, that is about 20% of the audience. Out of those 6, 3 were on the panel with him, so only roughly 10% of the audience. Most of the attendees do not have experience and were looking to see what the trends are and identify key players in the Chiplet space. 

There is no doubt we will start to see new players in this space, including startups looking for an exit (there was a session on the final day called Chiplets for Entrepreneurs in which a panel of “Industry experts” and incubator/VCs were giving advice on how to approach this market). Moving forward we will start to see a stronger push for standards and simplification which I believe will create commodity products which almost always results in consolidation. But we are still far away from this.

At Credo we are a leader in the IP and Chiplet business. We are in mass production and shipping Chiplets utilizing XSR and BoW (bunch of wire) interface and we have a roadmap that will continue to define us as a leader in this space. The adoption of standards will take time and the barriers to entry will remain high, allowing us to continue to be a pioneer in this space. 

We will have to wait for next year’s Summit to see the overall progress of the Chiplet business and what improvements and solutions are created in 2024. If you would like to learn more about our Chiplet and IP portfolio contact sales@credosemi.com. Thanks to the team at Yole for sharing their data with us.