As new manufacturing nodes come out, their deployment is increasingly expensive, due to high costs, a good part of the digital part of the hardware is being discontinued in the newer nodes. But the high cost means that hitherto unprecedented solutions have to be found.
CPU and GPU manufacturing, who makes it?
Many of our readers will already know this, but there is a difference between those who design a chip and those who manufacture it en masse. For example NVIDIA and AMD do not have their own factories and what they do is manufacture in a third party such as TSMC or Samsung. While companies like Intel have their own factory or foundry.
To make a comparison that helps you understand the situation, it is as if Intel were a publisher that had a series of authors of books on the payroll of which it edits them exclusively, while TSMC and Samsung act more like free printers that in exchange for renting their printing presses they publish the books that they ask for.
The different chip manufacturers or foundries compete by developing new manufacturing nodes and thereby attract potential customers who seek to be able to manufacture chips with more transistors, lower consumption and capable of reaching higher clock speeds. But, such a fight could come to an end.
The chiplet era, multiple chips to do one’s job
The existence of chiplets is not a secret, it is about dividing the functionality of one or more chips into several different ones in order to be able to increase production. The reason is that the more a chip grows in size, the more expensive it becomes to manufacture, since the number of potential failures grows because there is a larger area.
The great players in this endless theater play that is the hardware market have already announced that the future of several of their processor designs will go through the use of chiplets and especially GPUs. We have cases like Intel’s announced Intel Xe-HP, the rumored NVIDIA Hopper GPU, and AMD’s chiplet patent. All of them designs designed to be manufactured using chiplets. But not only the GPUs will be the ones that manufacturers will evolve the chiplets but also the CPUs, something that AMD has already done with its AMD Ryzen.
The fact of going from using a single monolithic chip to using several allows new types of configuration and transforms into changes in the supply chain when launching a processor, since we go from having one component to having several.
Different chips, nodes and manufacturers in a single CPU
One of the things that many people do not know is that not all chipmakers have the same strengths, since it may be that two different foundries with very similar nodes have specialized in very different fronts. For example, one manufacturer might give more importance to its manufacturing process to the density of the SRAM, while another prefers to give importance to other fields such as heat released.
Being forced to choose a single manufacturer for your CPU means taking away all the advantages and disadvantages of its manufacturing process. But, with the use of chiplets it is possible to use several foundries for the different manufacturing parts of a new processor. In the future for example we may see a new processor with parts from Intel, TSMC, Samsung and other foundries. Much like what AMD does today with chiplet-based AMD Ryzen.
We know that when it comes to CPU manufacturing such as Intel and TSMC are willing to collaborate, many talk about Intel manufacturing at TSMC, but those who talk about it are not thinking laterally. With chiplet-based designs around the corner, it is more than certain that we will see designs with pieces from different foundries and not just one.