Scientists, the industry, the stock market… – in short: everyone is getting excited about chiplets. These are individual modules that, for example, add several CPU cores, graphics units, or other components of the entire chip. Dr. Michael Töpper speaks to RealIZM (Blog of Fraunhofer IZM) about whether this new system architecture can really become the basic structure of the future industry, whether they are the answer to Moore’s law, and whether scientific progress can keep up its pace in these developments.
RealIZM: What is a chiplet?
MICHAEL TÖPPER: That is a very good question, because there is not a standard definition for a chiplet. A chiplet is one of the possible routes forward for semiconductor engineering. Because we are, once again, seeing what might be the end of Moore’s law. Moore’s law holds that device structures are getting smaller and smaller. Unfortunately, production costs are increasing so much that it is no longer commercially viable to go on into a sub-14 nanometre scale. 7 nm is coming into production quite soon, and if you are designing a new device on this scale, you have costs going into the hundreds of millions for each new chip generation. The idea is now to use different kinds of IPs, the intellectual property of the chip design, which can be used for certain functions. The different designs are already tested and can be put together like a puzzle, so you could use already existing parts and design only the important parts anew, for example the sub-7 nm. For instance, you take all of the analogue parts in an IC and add the functions you want to design new on a chip.
So essentially, a chiplet is not a fully functional stand-alone chip, but part of a chip that you can combine with others. Right now, we are using SoCs, which stands for Systems-on-Chip, containing all functions on one die. Another concept, which is a bit older, is the multichip module (MCM). In an MCM, you have packaged chips forming a modular system. But the chiplet is the new idea to move several functions out of the devices. One of the primary goals is to keep the costs down, because the costs are exploding now due to this extreme miniaturisation.
RealIZM: So the main goal is to simply keep the costs down?
MICHAEL TÖPPER: You want to have higher performance, and higher performance today is mainly driven by applications like artificial intelligence or autonomous driving. In this area, you need a lot of computing power, and you need an economical solution for that. You also hope to bring down power consumption, because power consumption is the big challenge for driverless cars. You simply do not have that power in a car, especially when you have an electric car.
RealIZM: And are chiplets the answer to Moore’s law?
MICHAEL TÖPPER: If you read Moore’s actual paper, it is interesting that he already explained in 1965 that “it may prove to be more economical to build a large system out of smaller functions”. Even Moore himself saw at that time already that there is an end to what came to be called Moore’s Law. By the way, people are talking about a “law” the whole time, but Moore’s law is not a law. It is a prediction.
RealIZM: Do you think it is possible to realize this prediction with chiplet technologies?
MICHAEL TÖPPER: Well, maybe we can. Chiplets might be one possibility. That is the reason for why the technology is one of the hottest topics at the moment in the semiconductor area.
RealIZM: How will chiplet influence the future? How will they change our lives?
MICHAEL TÖPPER: Our life has been already changed a lot due to miniaturization. And I think the next big thing will be artificial intelligence and autonomous driving, so we can say “the robots are coming” (laughs). But it is almost impossible to say what will happen in ten or fifteen years. Looking back to 20 years ago, nobody would have predicted the smartphone.
Look at what happened at Nokia. Nokia is a great company, but they failed totally because they did not see the potential of the smartphone. Nokia was the largest cellphone maker in the world, and now its only number nine or ten. They lost it, because they did not realize the possibilities of semiconductors and their role for system integration.
So, what’s coming up in 15 years? We cannot know. If you have an answer, I would invest in the company (laughs).
RealIZM: What do you think: How many years will it take to get this chiplet technology actual into application?
MICHAEL TÖPPER: There are already chiplets out there. It is all mainly driven by three or four companies. Intel is certainly one of them. They have a platform they call Stratix, which is a high-power semiconductor device that you can buy it already. And you can buy the technology from Tesla, where they are also working with chiplets in autonomous cars. AMD is working on the concept as well. And there is Samsung. These are the main companies driving this concept, putting together high-performance computing with extended memory.
RealIZM: And which problems can you see being solved by chiplets? You already mentioned high performance…
MICHAEL TÖPPER: I would say: low-power, high-performance computing. Right now, this is the main thrust, but you need a lot of intelligence for autonomous systems. You might have the systems, but they are bulky workstations that need a lot of power or that need a lot of space. Look at Volkswagen right now: They are testing autonomous cars (Golfs), but they have no space left in the trunk, because it is completely taken up by one massive computer. You have to miniaturize and lower the power consumption.
RealIZM: What is the current state of the art for this technology? Can you tell us more about these examples?
MICHAEL TÖPPER: Take the Stratix from Intel; They have a 10-nanometre FPGA connected to a high-bandwidth memory using a smart interconnection. They can go up to 112 gigabits/s, using this very high-power data transfer technology. This is where Fraunhofer IZM is coming into play, because we are working on interconnections for these devices, because high-density interconnection is one of the gatekeeper goals you simply have to achieve. The systems are running very fast, but you have to have ways to transfer the data. Now, the question is whether the data will be transferred electrically or optically. This is why we are also working on optical data transmission between the different components.
RealIZM: Are there any other examples in the market apart from Intel?
MICHAEL TÖPPER: As I mentioned, AMD and Tesla are working on this, and you know there is a large consortium in the US, run by the Defense Agency, the DARPA.
RealIZM: What is the difference between these chiplets and these concepts?
MICHAEL TÖPPER: One point of difference is the interconnection. Intel has what they call ‘embedded bridge components’. They achieve the interconnection between the components by using an embedded small silicon with high-density routing into organic substrate. The others are generally using silicon interposers, which is what we have been doing for eight years in Dresden and Berlin. One of the primary goals is a high-density routing substrate. It might happen that you have not only a passive substrate, but also some active components, for example voltage regulators or other passive components. As I mentioned, it may be electrical, but it might also be optical. By using the optical version, you could save energy, and it is very fast.
One of the biggest challenge at the moment is that we need to create standards. In the end, we will have chiplets from Intel, Samsung, and many others, so then we need standards to ensure that the chiplets fit together.
RealIZM: What are the goals or vision for IZM in this field?
MICHAEL TÖPPER: I think our vision relates to the interconnection substrate. Here at IZM, we have two technology lines: one is the wafer technology line, where we are working on high-density interconnections on wafer form, but we have also the panel line, which works with organic substrates. And this is where we want to increase the routing density.
So, we are going to planning for two micro lines that we might be able to use in large panels for the interconnection substrates.
RealIZM: And what can you see coming next?
MICHAEL TÖPPER: I think we’ll have to spend the next 10 years on chiplets. That is not something we can finish in just one or two years. We already have the first products, and we will see more around in about two or three years. But I think the main boom will come in six or seven years. That is also when we might see driverless cars on the road, which would be a great market for chiplets. You need a lot of computing power for these autonomous cars, because they have to manage a lot of data from their cameras, LIDAR systems, and so on, and somebody has to manage all that data on board.
The full version of this interview you can find under: https://blog.izm.fraunhofer.de/will-chiplets-revive-moores-law/
RealIZM is an online communication platform and corporate blog of the Fraunhofer IZM, dedicated to sharing our R&D ideas and technological solutions on application-oriented topics and to providing insights into breakthrough technologies in the field of microelectronics. The blog represents a profound scientific forum for communicating with electronics enthusiasts and making ideas come to life.