When Can I Buy A Chiplet?

One year ago, Semiconductor Engineering conducted its first roundtable to find out the true state of the industry for chiplets. At that event, it was stated that no chiplet had ever been reused in a design for which it was not initially intended. How much has changed over the past year? Returning from last year were Mark Kuemerle, vice president of technology for Marvell; Letizia Giuliano, vice president for product marketing and management at Alphawave Semi; Hee-Soo Lee, HSD segment lead for Keysight; Mick Posner, senior product group director for Cadence’s Compute Solutions Group; and Rob Kruger, product management director for Synopsys’ multi-die strategy solutions group. What follows are excerpts from a roundtable discussion held at this year’s Design Automation Conference.

SE: At this time last year, a similar panel of experts said that chiplets faced many hurdles and that reuse has not yet been accomplished. There were a lot of problems to be solved, and many obstacles that had to be overcome. It was also clear that not all of the obstacles were even identified, and there was certainly no consensus about how they should be solved. How much progress has been made in the past year and are we any clearer about the path forward? Why can’t I buy a chiplet today?

Kuemerle: I’ve been a fan of chiplets for many years, essentially since the beginning of the movement. When Marvell started getting involved in this, we envisioned this grand idea where we’d be able to source chiplets on an open market to do any given function. We would be able to define interfaces and protocol layers that would allow them to communicate together. We were completely wrong. It was important that we had this bold drive, because it forced us to go and figure out how to start the ball rolling. But what we’ve learned, by implementing many systems with multiple chips and chiplets, is that with a 2.5D integration strategy, we plan designs down to the micron from an alignment point of view. Beyond that, we have to plan so that the next layer up can communicate effectively between the chiplets in the most efficient manner. If you really want to optimize a system, they tend to be bespoke systems. However, I do want to add one change that I’ve seen over the past couple of years. These are the years when I say that chiplets are the tail that is wagging the dog. The idea is that chiplets are defining the rest of the system. If we have a standard chiplet that does a standard function that multiple people want, that system can get defined around that chiplet. We can achieve these things where we match everybody up to the micron, because the rest of the system is then built around it. We’re at this incremental baby step where we’re starting to get to a point where chiplets can be used for multiple projects. But it doesn’t mean that we can just take two chiplets and plug them together. While we can have independent chiplets, the glue has to be very bespoke to make those chiplets work together. We’re getting further, but we’re not at this open community, open market idea yet.

Giuliano: You asked when you could buy a chiplet. Actually, we’re showing a chiplet on the floor at the Design Automation Conference. Our own standard chiplet. We asked the industry if there was a way to disaggregate the system today, and the easy way is to segregate the I/Os. This is because they are so standardized. Connectivity is ubiquitous, and almost every device has the same interfaces that allow it to connect to the rest of the world — Ethernet, PCI Express. It is a natural way to capture the system and standardize how we do I/O to the chip. We did that with a large investment from our side, but it was also a response to the marketing need for dividing the system into chiplets. You are correct that we design around the chiplets that are available. Once people see that a chiplet is available and decide to use it, they start to organize the system based on what you have. Now we are creating momentum around that I/O chiplet. I accept that there are some issues. Definitely. There is a lot of standardization missing. The world has not yet standardized the design of the interface. We are missing many layers in the standardization. We do have something around UCIe, but it is missing some important pieces. The protocol layer is not standardized. That is a big gap. And we also have other layers in terms of form factor standardization. But we are the pioneers. We are trying to make sure that we solve the problems, and we have started building solutions so that people can build around them.

SE: Is that good enough for a company like Marvell to buy your chiplet?

Kuemerle: It depends. You certainly can get to a first degree where you have a chiplet that may be available. You talk with a provider and you figure out whether that fits into the puzzle that you’re trying to solve. One of the things we struggle with is that scaling technology is slowing down, but our customers aren’t slowing down in their need to keep getting as much out of technology as they can. The floorplans are growing bigger and bigger. The larger those floor plans get, the more complicated they get, and the harder it is to fit a given chiplet that you may find on the open market into that increasingly complex device. We find a lot of times that you really have to make a bespoke chiplet for some of the most complex applications. If you’re making something a little bit simpler, something that just needs 64 lanes of 112 gig SerDes, that might make a lot more sense. Perhaps you could purchase something from the open market when you have more degrees of freedom from an architectural point of view to put those pieces together. If you’re building a really big LEGO house, you probably want the purpose-made LEGO blocks that make it look nice and fit together perfectly. Whereas, if you just want to put some wheels together and make a thing that looks like a car, you can probably do it. There are different levels of integration, and that makes a difference about where you can use them.

Giuliano: There is no one solution that fits all. It needs to be customized for a specific application or use case.

Posner: Two years ago, I made a prediction. I said that a chiplet marketplace was 10 years away. By that reckoning, we have 8 years to go. It’s not like everything has stood still. There has been progress, and I believe it’s coming in a couple of areas. First of all, let’s distinguish it.  Multi-die has become a lot more mature. It is understood. A couple of years ago, it was still a visionary sell. When your design increases in scale, can you scale the package?  No, but that is tried and true now. The standards around it have started to mature. UCIe has matured. Arm introduced the chiplet system architecture, and that solves a little bit more of the puzzle. And then, ecosystems are coming up. One example is the imec automotive chiplet program (ACP). They look at the problem from a slightly different angle, and they solve another part of the puzzle. The realization of standards, processes, and manufacturing have matured to a point where I now believe there will be a chiplet marketplace in the future. We are moving in the right direction.

Fig.1: Chiplets: New way to build SoC. Source: Cadence

SE: Do you still think it will take another eight years?

Posner: Yes, for production. The other thing that we have to deal with is terminology. The term chiplet is an overloaded function, and right from the start I always have to explain to people what we mean. Finally, people have started to understand that there are bespoke systems where people are doing a multi-die design, and they are doing it because they are forced to do it. And then there are chiplets, where you’re expected to take a piece of IP from a third party and integrate that into your system. It has become very application/ ecosystem-centric. Consider server AI. There are specific applications — I/O disaggregation, memory disaggregation. In automotive, you’ve got compute disaggregated from I/O, disaggregated from AI. There seems to be a smaller leap to a chiplet marketplace if you constrain it to an application specific area, rather than just broadly saying there’s going to be a marketplace.

Kruger: HBM is probably the shining star for the chiplet marketplace, but people want more. There’s always going to be room for customization and the desire to have something that goes beyond the standard, and is different than the standard. That’s always going to happen with the biggest players. Then you have a chiplet that has a pre-defined shape, and that means you have to design the rest of the system to match. You might not have that freedom to move that UCIe, or maybe you have some other constraint on how the PHYs are stacked. That might happen because of the form factor and what’s sitting next to it. There are many aspects that have to be considered at the system level. Eventually there will be off-the-shelf chiplets. But there’s also this need for customization. There’s going to be a balance. Then the question is, for commercial chiplets, is there a big enough market where people are willing to make some concessions, either on size or power?

Lee: Generally speaking, I agree. A standard application is very important, but you also need the ecosystem, because it’s only one part of the design. Where are those chiplets coming from? A lot of IDMs are making chiplets, but when I looked at the design itself, it may be in a chiplet style, but it’s not completely following the standard because they have their own needs. They have their own fabrication facilities, so they can design the other structures in a similar manner, but not exactly following the standard. This does not enable plug-and-play. This is where I see a gap. It takes time. Once you move outside of a custom-driven approach to being standard-driven, including UCIe and Bunch of Wires, then we will see a lot more opportunities. But based on the trend in the market and technologies, it could be quite a long time. I’m not saying how many years, but in order for the market to mature, it would take several years of investment. When looking at packaging in the past, and then considering the emergence of chiplets, it is important to think about how this is different from system-in-package. There’s a lot of packaging technology that gets mixed in with the system within a single package. The technology is more mature and ready to go for the implementation of chiplets, compared to the situation 10 years ago. It is possible that we can make this real now. But still, how are we going to get rid of those boundaries? How does the industry migrate from wanting everything custom into a more standardized way of integration? This is where the industry needs to focus.

In the next part of this discussion series, the experts discuss the transition from needing chiplets, to wanting chiplets, and the emergence of ecosystems.