集微网消息，在往期的集微访谈栏目中，爱集微有幸采访了The Linley Group总裁兼首席分析师Linley Gwennap，关于Chiplets，包括供应链会如何调整以适应社会对Chiplets的新需求、机遇与挑战、制造成本、封装生态等方面提出了一系列问题，并收到了十分有启发的答复。
问：近期，美国颁布的《芯片法案》里提到了Chiplets平台的建设，中国也越来越关注Chiplets领域。现在，你认为该法案将如何影响或重塑Chiplets的发展？
答：在过去的几年里，人们对Chiplets技术的兴趣和投资不断增加，这一技术正风靡全球。随着《芯片法案》的通过，我们将看到更多的投资涌入这一技术。政府投资将有助于加速Chiplets及其产品的发展。有了政府的支持，上升趋势会更加明显。但在技术或战略方面，不会受到影响。
技术和策略发展得很成熟。像英特尔和AMD这样的大公司已经采用并且开创了几种不同的Chiplets技术，台积电也不断开发封装技术，在Chiplets技术方面取得了相当快的进展。因此，在这一点上，想法已经很成熟。政府资助可能不是百分百需要，但相比公司自己投资，肯定会加速推动技术发展。
问：像Chiplets等新技术仍然缺乏合适的设计工具、测试和封装解决方案等。您认为目前供应链会如何调整以适应社会对Chiplets的新需求？
答：这是个好问题。尽管我们已经建立了一些设计和制造Chiplets的技术，但关于如何测试新系统，是否有一些基本想法？
现在的情况是不止一个芯片，而是几个芯片一起工作。为了能够处理这些制造技术，设计工具必须进行修改和升级。工具公司意识到了这一点，注意到这种趋势正不断发展。因此，他们正不断研究技术，并对工具进行补充更新。但现在，拥有更先进的设计能力且可以独立制造此类Chiplets的公司屈指可数。但我认为，随着时间的推移，未来几年，将有更多的公司踏入Chiplets设计领域，对标准工具，代工厂和封装供应商提高的标准制造技术的需求大大上升。因此，整个工具、测试和封装技术的基础设施的建立需要更长的时间。
目前，第一波Chiplets产品来自大型供应商，而未来3-5年，一旦工具和基础设施以及供应链更加完善，我们将看到能够利用同样的Chiplets技术的公司会越来越多，
问：你的这个回答与第三个问题有些相关，就是对于拥有不同市场份额、不同规模的公司而言，Chiplets带来了哪些机会或挑战？
答：Chiplets的设计过程和制造过程更复杂。在面对建立单一芯片还是将其分成Chiplets的抉择时，总会有一些权衡。
跨进该领域的公司是研究这种新技术的先驱，它们正努力解决其中遇到的挑战。例如，如何让信号在Chiplets之间来回传递？有没有一些标准方法来做到这一点？如今相关标准仍未建立。如何测算这些信号的传播速度和能力？今天还没有出现好方法。理想状态是，我们应该能够从不同的公司获得Chiplets，并使用标准技术、标准接口和标准封装将它们连接起来，虽然现实离它还很远。提供芯片技术的供应商很少。例如，AMD有一些产品，它能独立设计的芯片很Chiplets，将其应用到工艺或系统中
他们可以控制Chiplets一起工作并在封装内相互配合。因此，这种模式正在拓展，允许更多公司建立Chiplets。然后其他公司则负责把它们整合在一起，这将需要做的工作更多，标准也更多。因此，Chiplets技术可能要在3年后才能实现。但我认为这确实是很多公司想要达到的目标，就像一个大公司授权建立arm核心和GPU核心，并将其与自己的技术相结合，共同研制单一芯片。所以对于Chiplets而言，你可以从一家公司获得带有CPU的Chiplets，从另一家公司获得图形加速器，然后建立自己的Chiplets，其中包含自己研究的特定技术。
你需要把所有这些东西拉到一个封装内，让它们配合工作。现在能省去很多麻烦，不用单独处理那些可以从第三方获得的组件，这就是我们想要实现的理想状态。但现在，我们还处在Chiplets产能的早期阶段，还需要很长时间。
问：在这一点上，遇到的挑战或机会可能会因国家而异，特别是在中国会有什么具体表现？
答：在中国，越来越多的公司正在进入芯片设计领域，他们不一定有那么大的知识产权库、没有自己的CPU和图形。
Chiplets技术和这种理想情况将使这些新公司更容易进入，例如，建立大数据中心处理器，不需要自己做完所有的事情。甚至你可以想象不同的公司专注不同领域，分别做 CPU、AI加速器和内存，然后这些Chiplets一起工作。我认为Chiplets是技术，可以帮助中国等国家更迅速地建立半导体基础设施，很多新公司聚集在一起工作，但如果是大型、有经验的公司，可以自己完成所有设计。考虑使用Chiplets的原因个不相同，但我认为，特别是对于这些正在中国等地方建立的新公司，Chiplets生态系统可以创造新的机会，加速将新产品推向市场的能力。
问：正如你所说的，从工具到制造，所有技术都开始加速发展。随着美国在芯片制造和设计工具上投入更多的资金，你认为Chiplets可能是中国公司的另一个出口吗？
答：Chiplets能够实现的不是依靠前沿技术设计一个大芯片，而是将它进一步分割，分割的部分并不是都需要采用前沿技术，可以单独将其中一些部件放在技术更适用的国家，如中国，并在当地生产制造。也许有的Chiplets必须使用先进的节点，有的能在其他地方制造或从外部公司购买，然后所有这些Chiplets再集成到一起，建立一个可以在中国等国家使用的系统。因此，它可能是一种方法，通过将一些Chiplets转移到现有的节点，并最大限度地减少对前沿技术的需求，来解决缺乏前沿制造的问题。
问：你去年五月好像发布了一篇文章，其中提到，Chiplets的一个挑战是建立生态系统。现在，你认为在这份生态系统地图上还缺少什么，哪些部分仍待建立？
答：我认为现在最缺少的是将Chiplets连接在一起的标准方式，而这正是我们真正需要的，让来自不同公司的Chiplets能够一起工作。自从我写了那篇文章后，有一个叫做UCI或通用Chiplets接口的新团体不断发展。许多大的芯片公司都参与进来了。我认为他们正在做的工作，以及初步建立的标准，将对创建这个在未来几年内每个人都会使用的接口有很大的帮助。
解决这个常见的接口问题方面取得了一些进展。但各公司需要时间来设计新的Chiplets以满足新兴标准。这就是为什么我认为仍然需要2或3年时间，才能实现从不同的公司购买芯片并使用这种新的UCI接口将它们连接在一起。目前。已经取得了一些进展，正在向这个目标靠近。对于公司而言，现在是关注Chiplets的最佳时间，能够利用这些新兴的标准和即将推出的新技术开始设计Chiplets。如果今天开始这项工作，在几年内，就可以把它全部拉到一起，设计出Chiplets，便可以向客户销售。
问：在UCI出现之前，各公司已经有各种互连接口，如Aib等。你认为所有这些公司都准备转到UCI，或者尝试保留自己的设计？
答：我认为对于那些自己设计Chiplets的公司来说，他们要设计所有Chiplets，并用在自己的产品中，继续使用自己的内部标准。但我认为，对于那些希望能够相互合作以整合来自不同地方的Chiplets的小公司来说， UCI标准的重要性更大，它鼓励多厂商的连接合作，使建立这些新组合成为可能。大公司可以坚持自己的标准，有个叫做ODSA的组织正试图推动很多其他标准，但我认为其中这些标准会被纳入UCI的考虑，这将是适用于那些想把Chiplets连接在一起的公司的真正的共同标准。
问：就技术和商业战略而言，独特的互连设计与你刚才提到的UCI或ODSA等标准相比，有哪些优点和缺点？
答：我认为互连本身是很常见的，这并不是特殊的技术。真正的问题是，如果你要把两个Chiplets结合在一起，两者必须统一种语言，没有优劣之分，只是两者需要统一，比如，每个人都可以同意用UCI作为连接Chiplets的方式，这样，问题就解决了。就技术能力而言，可能需要不同数量的带宽，需要不同的速度来发送信号，UCI团队正在定义接口，以便从技术角度满足不同的需求，只要有这样一个可扩展的接口，就可以满足所有人的需求。谢谢。
问：现在让我们稍微回到生态系统的话题上。你认为不同的国家会制定自己的生态系统建设计划吗？关注焦点可能会有所不同？
答：我们都不希望生态系统的建立受到地域的限制。如果我们有一个单一的标准，世界各地的公司都可以使用，真正鼓励创新，并允许全国各地的公司交换Chiplets，将来自不同公司和不同地域的Chiplets连接起来。如果在不同的地方有不同的标准，公司很难有自由交换的机会，人们也会忙乱于各种不同标准，这样同一个Chiplets只能在两个地方工作。我希望在这个层面上，可以有一个单一的标准，以允许在芯片上有更多的全球合作。
这个想法很积极，媒体应该推崇这种正面的想法。下一个问题是关于Chiplets的未来，考虑到所有不同的应用领域，您认为有哪些具体的应用领域在先进的Chiplets技术中受益最大？而其他领域可能不想急于采用Chiplets技术。
我们今天看到的Chiplets主要应用在昂贵的产品中。特别是在数据中心，例如，你需要非常强大的服务器处理器，希望在该服务器中内置人工智能加速器，现在需要的是将其分置在不同Chiplets上而非堆放在一个芯片上，这样生产工序简单一些，但问题是，Chiplets技术本身可能会增加一些成本，因为需要在一个封装中放置多个Chiplets，程序很精准，以便它们可以直接连接并相互配合。对于这些数据中心产品来说，这一点额外的成本并不是什么大问题。提到PC处理器时，我们开始看到一些Chiplets技术，它们对成本更敏感一些，但我认为如果以足够简单的方式进行，Chiplets仍然是有意义的。我们不会很快在智能手机或其他类型的消费设备中看到Chiplets，因为这些产品非常注重最小化成本。因此，Chiplets在该类产品中的作用不大，因为增加的成本可能比真正节省的要多。
而且Chiplets能够发挥最大价值的地方是先进技术，因为各家公司都在向五纳米晶体管和三纳米晶体管进军。这些晶体管成本昂贵，不想用这些昂贵的晶体管构建整个芯片，就只想把对速度需求最大的那一小块拿出来，用最小的晶体管来构建，然后在处理器的其他组件上使用较便宜的技术。Chiplets技术对于旧节点中的芯片意义不大。也许一些旧的消费设备使用了很多旧芯片，那么它们不一定需要转移到这种技术。我认为Chiplets技术不会出现在各个领域，也不是每个芯片都会被分割成Chiplets。最大的机会是这些技术领先的芯片，特别是那些昂贵的大芯片，才能看到最大的价值，将它们分割成Chiplets能获得最大好处。
问： 一些分析师对其在汽车行业的Chiplets发展非常乐观。你同意这种说法吗？
答：在如今的汽车中，有数百个芯片，其中大多数是建立在旧技术上的，它们做的事情相当简单。而这些芯片，我认为，不一定需要Chiplets。展望未来，我们看到一些更领先的芯片进入汽车，特别是自动驾驶汽车，显然需要非常复杂的处理器，以能够驱动整个汽车，甚至一些半自动系统会帮助汽车正常行驶，在紧急情况下能够自动预警，这些芯片也是相当复杂。在高端汽车中可能有一两个芯片可以从Chiplets技术中受益，它不是适用于汽车中所有简单和通用的芯片。
问：下一个问题是关于成本。工业界希望从单片式芯片转移到Chiplets的原因之一是，Chiplets的成本通常较低。但我们也看到有报告说，将所有的技术或设备从单片机转移到Chiplets结构是非常昂贵的。那么你对成本问题有什么看法？
答：这是一个复杂的问题。正如我所说，当你采用Chiplets设计时，会有一些额外的成本，因为封装两个或四个芯片比一个芯片更昂贵。现在考虑到成本节约，这就是为什么人们说Chiplets可以省钱，因为不必把所有的电路都放在这些非常昂贵的前沿晶体管中，可以退而求其次，把一些东西放在旧的工艺中。我做了一项研究，得出的结论是，对于大于300平方毫米的芯片来说，有一个交叉点，Chiplets技术开始节约成本，当使用非常大的芯片时，就是最大的芯片，六七百平方毫米，Chiplets技术就会有很大的节省，300平方毫米以下时，至少在今天的技术基础上，成本开始向另一个方向倾斜。
增加的成本比你节省的要多，这个交叉点可能会随着时间的推移而发生改变。但现在，300平方毫米的范围很广，包括很多数据中心的芯片，但PC芯片往往在200平方毫米的范围内，智能手机芯片主要在100平方毫米范围内。
这就是我觉得对于其中的一些芯片来说，想要立即转移到Chiplets技术上不太合适。但是，一旦我们的生态系统更广或Chiplets技术背后的数量增加，成本应该可以下降，交叉点将下降，使用Chiplets技术的芯片范围会增加，它将随着时间而增长，从一些更高端的应用开始，逐步进入低成本的应用，可能需要5-10年的时间才能真正进入低成本。
以下是采访原文（英文）：
Q:China"s being really paying a lot attention to chiplet, partially because the chips act are specific mention the building of chiplet platform. Now, how do you think chips act would affect or maybe reshape the development of Chiplets?
A:I think that there"s already been a lot of interest and investment in chiplet technology over the past few years. So it"s a technology that"s really catching on. But I think with the passage of the chips act, then we will see more investment. And so I think that government investment will help to accelerate the development of Chiplets and chiplet products, so that the trend will move even faster now with backing from the government.But in terms of technology or strategy, there wouldn"t be any difference.
I think that the technologies and the strategies are already fairly well developed. We"ve seen big companies like intel and AMD adopting chiplet technology already. They"ve pioneered several different technologies for Chiplets, TSMC has been developing its own packaging technology as well. So those companies are already moving forward pretty quickly with the chiplet technology. So I think the general ideas are pretty well established at this point. And the government funding may not be totally necessary, but it will certainly help to move things along a little more quickly than just the companies investing on their own.
Q:Thank you. With a new technology like chiplet industry is still lacking of proper designing tools, test and packaging solutions, etcetera. Now, how do you think the supply chain would adjust to the new needs of Chiplets?
A:Yeah, that"s a good question. And even though we have some of these techniques established to create the Chiplets and manufacture them, some of the basic ideas of how do you test a system.
Now, it"s not just one chip, but it"s more than one chip, several chips working together. And the design tools have to be modified and upgraded to be able to handle these new manufacturing techniques. The tools companies are aware of this. They"ve been seeing this trend for some time now. So they"re already working on techniques and additions to their tools to make this easier. But right now, it"s only few companies that have more advanced design capability that they can make these kind of chiplet designs. But the idea is, over time, the next few years, that the more companies will be wanting to make chiplet designs. Those companies are going to need standard tools. They"re going to need standard manufacturing technologies from the foundries and the packaging vendors. And so it will take a little more time for that whole infrastructure of tools and testing and packaging technologies to catch up.
So I think right now, we"re seeing sort of first wave of chiplet products coming from the biggest vendors.
And then over the next 3 to 5 years, we"ll see more companies and smaller companies able to take advantage of the same chiplet technologies once the tools and the infrastructure and the supply chain is better established.
Q:Thank you.So I think you touch on a little bit of my third question is about for different companies, with different sides and maybe various market shares, what new opportunities or challenges does the chiplet bring?
A:I think that the chiplet does definitely complicate of the design process. It complicates the manufacturing process. And so there"s always tradeoffs in deciding whether you want to build a single chip or whether you want to divide it up into Chiplets.
And so the companies today are kind of pioneering this technology, are trying to figure out ways to get around some of these challenges. For example, how do you get the signals to go back and forth between the Chiplets? Is there some standard way to do that? There"s no real standard today. How do you measure the speed and test the capability of those signals that are going back and forth? There"s no really good way to do that today. And in an ideal world, we should be able to get Chiplets from different companies and combine them together using standard technology, standard interfaces, standard packaging. But we"re really not very close to that today. So the chiplet technologies that we"re seeing are really coming from just a single vendor. For example, AMD has some products where they take chips that they"ve designed, Chiplets that they"ve designed and build them into a process or a system.
They can control all of the different aspects of how those Chiplets work together and how they"re going to coordinate in the package. So trying to extend that model, to allow multiple companies to build the Chiplets. And then someone else, perhaps to integrate them all together. That is going to require more work and more standards in order to accomplish that and. So that sort of multi chiplet ideal is probably going to be 3 years down the road. But I think that"s really where a lot of companies want to get to where the same way that a bigger company can license an arm core and license a GPU core and and put it all together with their own technology on a single chip. The idea is with Chiplets, you could get a chiplet that has armed CPUs on it from one company, maybe get a graphics accelerator from another company, and then maybe build your own chiplet that just has some specific technology that you"ve designed.
You should be able to pull all that together in a package and everything works together. And now you saved yourself a lot of trouble of having to deal with these other components that you"re just getting from third party. So and that"s kind of the ideal that we"re trying to get to. But right now, we"re just in very early phases of the chiplet capability and is gonna take time to build out to that ideal situation.
Q:Great. On that note, do you think those challenges or opportunities may be vary by country, especially how would that look in China specifically?
A:Yeah, I think that in China, there"s a lot of different companies and a lot of the new companies that are getting into the chip design business. They don"t necessarily have as a big library of intellectual property. They don"t have maybe their own CPU. They don"t have their own graphics.
And so the chiplet technology and this kind of ideal scenario, I think, would make it easier for these Chinese companies that are new to get into, say, building big data center processors, because they wouldn"t have to do everything themselves.
And then even you could imagine different companies, one company makes the CPU, another company makes an AI accelerator, another company makes the memory. And then those Chiplets all work together. So I think chiplet is a technology that could be used to build more quickly semiconductor infrastructure in a country like China, where there"s just a lot of new companies coming together, I think, where you have big experience companies that have the capability to design everything themselves. There"re some different reasons why they would want to use Chiplets, but I think particularly for these new companies that are emerging in China and other places, the chiplet ecosystem could create new opportunities to accelerate their ability to get new products to market.
Q:So as you"re saying that everything is from the tools to manufacturing of the things kind of starting to boating right now. So also, as US has been putting more bands on the chip making and designing tools we already have now, do you think maybe chiplet is another exit for Chinese companies?
A:One thing that a chiplet enables is rather than designing a big chip on a leading-edge technology, you can divide it into smaller pieces, and not all of those pieces need to be on the leading-edge technologies. If you can take some of those pieces and put them on over technologies that are more widely available in China, then that would enable those components to be manufactured locally. Maybe one of the Chiplets has to be in a leading-edge node, and maybe that one could be manufactured somewhere else, or maybe it could be purchased from an outside company and then integrated in with the rest of the Chiplets to build a system that could be used in China and elsewhere.
So it could be a way to get around the lack of leading edge manufacturing by moving some of the Chiplets into the existing nodes and minimizing the amount of leading edge technology that you need.
I"m glad you mentioned the building of ecosystem before. I think in the report you published, maybe I think last year in May, you also mentioned that one of the challenges for Chiplets is to building the ecosystem.Now, what do you think is still missing on the map to accomplish the ecosystem and which parts are or like still waiting to be established or may be polished.
I think that the big thing that"s missing right now is a standard way of connecting the Chiplets together. And that"s what we really need to have Chiplets from different companies all work together. So the very recently, in fact, since I wrote that article, there"s been a new group come together called UCI or universal chiplet interface. And many big chip companies are involved in the UCI group. I think that the work that they"re doing and they already have some preliminary standards will go a long way toward creating this interface that everybody will be using in the next few years.
So we"re starting to see some progress in solving this common interface problem. But it will take time for companies to design new Chiplets to meet the emerging standards. That"s why I think it"s still gonna be 2 or 3 years before you could go out and start buying chips from different companies and plugging them together using this new UCI interface. But we"re definitely making progress and getting to the point where that"s going to happen. So I think you know as a company, it"s a great time to be looking at Chiplets to start designing Chiplets to be able to take advantage of some of these emerging standards and new technologies that are coming out. And if you start that work today in a couple of years, you could pull it all together and have a chiplet design that you could and start selling to customers.
Q:So before this UCI coming out, the companies already had various interconnect interfaces, like AIB etc. Now, do you think all these companies are ready to transfer to UCI or would they even try to keep their own designs?
A:I think for companies that are designing Chiplets themselves, and they"re gonna design all their Chiplets and keep them in their own products. They can continue using their own internal standards. But I think for smaller companies that want to be able to work with each other to integrate Chiplets from different places. That"s where this UCI standard will be important to encourage this kind of multi-vendor connections and make things possible to build these new combinations. I think that the bigger companies can stick with their own standards.
But a lot of the other standards, there"s another group called odsa that was trying to promote a chiplet standard, but a lot of those standards, I think, are gonna go away and be subsumed into this UCI effort. And that will be really the common standard for companies that want to plug their Chiplets together.
Q: So in terms of maybe both technology and business strategy, just generally, what are some pros and cons does the unique interconnect design have comparing with a standard like UCI or odsa that you just mentioned?
Yeah, I think that the interconnect itself is pretty common. It"s not that there"s any special technology there. It"s just a question of if you"re gonna plug two Chiplets together, they have to agree on what language they"re speaking, right? And It"s not that one language is better than another, but you have to agree. So I think that"s where you know, everybody can agree on UCI as a way to connect the Chiplets together. Then that should be able to solve the problem. I think in terms of technical capabilities, there"s a different amount of bandwidth that you might need. There"s different speeds that you might send your signal, but the UCI group is defining their interface so that it can scale up and scale down to meet different needs from a technical standpoint. So I think as long as you have a scalable interface like that, it can really pretty much serve everybody"s needs. Thank you.
Q: Now let"s move back to the ecosystem topic a little bit. So do you think different countries maybe will make their own ecosystem building plans? And maybe the focal points will vary?
A:I hope that we don"t see things become geographically are restricted. If we have a single standard that companies around the world can use, it will really encourage innovation and allow companies in different locations to exchange Chiplets, to plug Chiplets together from different companies and different geographies. If we end up with different standards in different places, it just will make it very hard for companies to have that sort of free interchange. And then people have to do create effort from one standard to another. Just so we can have the same chiplet working in two different places. And I hope that we can, at that level, have a single standard to allow a more global cooperation on Chiplets.
That"s definitely some optimistic positive thinking here. I think we like that on the media. So my next question is about maybe the future of the chiplet, but also considering all the different application areas, I would say. So what are some specific application areas do you think would benefit most from an advanced chiplet approach? Whereas others may not wanna rush to adopt Chiplets so soon.
Where we"re seeing Chiplets today is in the more expensive products. So particularly in the data center, for example, you need a very powerful server processor. And then you might want to have a very powerful AI accelerator built into that server. And instead of trying to put all of that on one chip, the ideas that then break it down into several smaller chips, which are easier to manufacture. Now, the problem is that the chiplet technology itself can add some cost in terms of the technology required to put more than one chip in a package and do that very precisely so that they can connect directly and talk to one another. And so for these data center products, that little bit of extra costs is not that much of a problem. As you look at pc processors, we"re starting to see some chiplet technology. They"re a little bit more cost sensitive, but I think the chiplet still can make sense if it"s done in a simple enough fashion. I don"t think we"re going to see Chiplets in smartphones anytime soon or other sorts of consumer gear where there"s very much focus on minimizing the cost. So the Chiplets don"t really work as well there, because there is probably more cost being added than you"re really saving.
And it"s also the Chiplets where we"re seeing the most push for Chiplets is kind of the leading edge, right? Because the companies are going to five nanometer transistors and three nanometer transistors. Those transistors are so expensive. You don"t want to build the whole chip using those expensive transistors. You just wanna take the little piece that needs to go as fast as possible and build those in the smallest transistors. And then you use an older technology, less expensive technology for the other components in the processors. So the chiplet technology doesn"t really make much sense for chips are already in older nodes. Maybe some of the older consumer gear cars use a lot of older chips, so they don"t necessarily need to move to this technology.So I don"t think Chiplets are just gonna be everywhere. I don"t think every chip is gonna suddenly become divided up into Chiplets, but where we see the most opportunity is these leading edge chips, and particularly the big chips of the expensive chips that are the ones where we"re seeing the most value, the most benefit to dividing them up into the smaller Chiplets.
Q: So some of the analysts are actually really optimistic about its Chiplets development in automotive industry. So I"m taking that you agree with that visionary？
A: In an automobile today, there"s hundreds of chips, and most of them are built on older technology, and they do fairly simple things. And those are the chips, I don"t think, necessarily need Chiplets. Now, it is true that particularly looking forward that we see some more leading-edge chips going into vehicles, particularly autonomous vehicles, obviously need very complex processors to be able to drive the entire car and even some of the semi-autonomous systems that help the car stay in the right lane or will break in an emergency situation, those chips are also fairly complicated. So there"s probably one or two chips in a high-end car that could benefit from the chiplet technology, not necessarily all of the fairly simple and generic chips in the car.
Q: My next question is about the costs. So one of the reasons that the industry wanna transfer from a monolithic chip to a chiplet is because Chiplets are generally less expensive. But we are also seeing reports saying it is very expensive to transfer all the technology or equipment from a monolithic one to chiplet structure. So what is your take on the cost issue?
A:It"s a complex issue. As I said, there is some added cost when you go to a chiplet design, because it is more expensive to package two chips or four chips than one chip. There is some cost addition there. Now there is also some cost savings, which is why people say Chiplets will save money, because you don"t have to put all of the circuits in these very expensive leading edge transistors that you can back off and put some things in an older process. You have to kind of balance off the addition and the subtraction. I"ve done a study. And what I came up with is that there is a crossover point for chips that are bigger than about 300 square millimeters, is where the chiplet technology start saving money. And when you get into very large chips, I mean some of the largest chips, six hundred seven hundred square millimeters, then there"s a big savings. As you go down below 300 square millimeters, at least on the basis of today"s technologies, the cost starts to tip more in the other direction.
And now you"re adding more costs than you"re taking away. So that crossover point will probably change over time. But right now, the 300 square millimeters pretty cover, certainly a lot of the data center chips. But pc chips tend to be in the 200 square millimeter range. Smartphone chips tend to be in the 100 square millimeter range. That"s where I think it"s going to be less common for some of those chips to want to move over to chiplet technology right away. But at once we get more ecosystem or a volume behind the chiplet technology, the cost should come down. The crossover point will come down, and we"ll see more adoption. So it is something that is going to grow over time. And we"ll see it start off in some of these more high-end applications, trickle down into lower cost applications that could take 5 or 10 years to really get down into the lower cost points.
