Why AMD Is Paying $35 Billion To Buy Xilinx

Most of the stories about AMD’s plans to acquire Xilinx XLNX Inc. for $35 billion focus on the consolidation of the U.S. semiconductor industry, following recent deals for ARM Holdings and Maxim Integrated Products MXIM . But this transaction is interesting because it says a lot about where computer architectures are going, and what Xilinx could mean to AMD’s customers and its product roadmap.

Xylinx is a leader in field-programmable gate arrays (FPGAs), a type of chip that has various resources on it such as programmable logic blocks and reconfigurable interconnects that allow different parts to be wired together. This is very different from the microprocessors from Intel INTC and AMD that we are more familiar with that go into PCs or datacenter servers. In the latter case, the chips have a specific instruction set architecture, and they do certain things – they run operating systems like Windows or Linux, and you program them with software to do different things. Once the chips come out of the factory (or “fab”), they don’t change. They are soldered on to circuit boards, and the only way to make them do different things is by changing the software that they run. Microprocessors have been so successful over the last 50 years because when they are produced inexpensively in huge volumes, you can make them do so many things just with software.

FPGAs are different. After they come from the fab and are packaged, you can change how the pieces are connected inside, making a unique chip. The roots of this approach traces back to programmable read-only memories from the early 1980s. The idea then was that if I present a chip with some combination of inputs, I want it to generate some set of outputs. That is what the goal of digital logic after all, whether I do it with logic gates on a chip, or a microprocessor running software.

Historically FPGAs have been slow performance-wise and relatively expensive, but companies like Xilinx have brought the latest chips to amazing heights. Today you can build custom devices with FPGAs that do all kinds of unique things, and you don’t have to have the large volumes behind Intel’s or AMD’s microprocessor sales to get the cost down. That “late customization” gives product designers a huge amount of flexibility to rapidly evolve their products without having to wait for the long development cycle of a new microprocessor.

This is why FPGAs are being used in many new bleeding-edge applications in machine learning, communications equipment, and a wide range of other devices.

The Xilinx acquisition points to a recognition that there is going to be more fragmentation in the microprocessor market of the future, with FPGA technology relieving many product designers from the constraints imposed by standard products. Intel purchased Xilinx competitor Altera ALTR back in 2015. For AMD, this gives them a level of insurance that they can remain in those customers’ roadmaps.