We picked a fight on Twitter this week. The proximate news was the announcement that Leendert van Doorn had joined Qualcomm. Our response is below:
Van Doorn is a highly respected computer scientist and engineer. He has done a lot of important work and thinking about the future compute needs of data centers, and was seen as a long-time proponent of Arm in the data center. This coupled with the fact that Qualcomm acquired Server CPU aspirant Nuvia last year, re-sparked speculation that Qualcomm was going to take a stab at making server CPUs, again. For a variety of reasons, some of which we cannot detail, we think that is unlikely.
So what is going on here?
In all the back and forth in the fight between Arm and x86 for data center CPU share, the industry tends to skip over a more important change in the space. Put simply, the old world in which data centers were giant warehouses filled with CPUs is gone. It is being replaced with warehouses full of a whole variety of chips - GPUs, AI accelerators, FPGAs and some CPUs. We could chalk this up to the slowing of Moore's Law, or the changing requirements of hyperscaler compute applications, and both would be accurate.
If you think about this bigger picture, as professionals like van Doorn are very good at doing, companies that operate data centers need to make some big changes in how they approach their silicon needs. This has led many of them to start designing their own chips, a topic we have touched on frequently. But even the biggest companies can struggle with this, as Facebook has demonstrated recently. And with a few exceptions, even the most capable companies need help in taking a chip from digital design through manufacturing to a production part.
Based on all this, our guess is that Qualcomm is looking to make friends with the big data center builders, but this will not entail productizing a general purpose CPU. Instead, they will help these companies design their chips internally. At its most basic, this could just mean handling the back-end of the chip design process, the interface between the designers and the foundries. This is an important role, it lets Qualcomm play off its established foundry relationships and operational scale. That being said, this is not a great business for merchant chip companies, margins are below chip and corporate margins typically, because at heart it is just a service business (i.e. it does not scale). So we think there is another model. The trend for many companies is to utilize systems on a chip (SoC) - that is a chip that combines functions of other discrete chips. So an SoC may incorporate CPU, GPU and AI cores, the Apple M and A series are good examples of this. Sometimes in these chips, the designer wants to design everything themselves, but it is also common to license some components from third parties. Memory companies, for instance, license the intellectual property (IP, their designs) to other chip makers.
Qualcomm's CEO has said their approach to the data center will be opportunistic, so if one of their customers is designing an SoC, Qualcomm could license some combination of processor and AI accelerator cores for incorporation into those internally designed chips. This has the benefit of being a healthy margin product, and Qualcomm is no stranger to IP licensing. Qualcomm can also build off relationships it already has with some of the hyperscalers - note the decade they spent porting Windows to Arm, or their work with Meta on AR chips. The other advantage here is that Qualcomm does not need to build a massive software team, since presumably the customer will handle that side of things.
Over time, we suspect that many of the hyperscalers will choose to focus their chip design teams onto a narrower subset of areas that offer the most strategic advantage. Meaning they will not want to waste time re-inventing the wheel for designs that they can readily license from companies like Qualcomm. This does not mean Qualcomm is going to launch whole new product lines, we see this is an add-on to existing relationships. This is just an incremental advance for a business model that Broadcomm and Marvell have been doing for many years.
No comments:
Post a Comment