The market for Compute Express Link (CXL) products will enjoy rapid growth to nearly $3.4 billion by 2028 as servers come to market with the technology.
That’s the conclusion of a new report from semiconductor market research firm Objective Analysis, which finds that CXL technology will gain popularity as computing systems are configured to take advantage of its many features and customers realize what it can do.
CXL is a CPU-to-device and CPU-to-memory interconnect that allows a processor in one server to speak directly to the memory in another physical server. It also enables any device to interact with the memory contents of the CPU. Communication is done through the PCI Express bus.
CXL attempts to address a number of memory issues, starting with available memory in a server. For example, a server under heavy load could run out of memory. If that happens, the server could connect through the CXL interface to another server under low load and borrow its unused memory.
Another challenge CXL solves is the so-called memory wall problem. As CPUs have developed, it has become a struggle to keep all of the cores fed with data. CPUs have gotten faster than memory, and the number of memory buses has also not kept up. A CPU with 64 cores might have only 12 memory buses between them, which means the cores have to share memory channels, which means latency.
“What CXL allows you to do is to add more memory channels onto the processor, without having to have just a giant number of pins on the processor for memory busses, and without having to consume the kind of power that all of those memory busses would [require], because a memory bus consumes a huge chunk of the power inside the processor,” said Jim Handy, president of Objective Analysis and author of the report.
The $3.4 billion may not seem like an impressive figure in this industry, but CXL chips average around $100. As CXL controllers find their way into one server vendor after another, the technology becomes widely available through increasing ubiquity.
The four major server CPU vendors – Intel, AMD, Nvidia, and Arm – are all incorporating CXL support in their processors going forward, and there are a number of controller makers that have licensed CXL, including Rambus, Marvell Technology and Samsung.
CXL will find its primary use in AI. AI is all about moving data, and if it can be done faster and directly between chips and memory, it will be used.
“If it’s all about keeping memory cores fed, well, then that’s AI’s play,” said Handy. “But it’s also going to benefit even databases and older technologies like that.”
Source:: Network World