CXL: A New Memory High-Speed Interconnect Fabric

The field of computing has been advancing rapidly in recent years, with new technologies emerging that are changing the way we process and store data. One of the most exciting new technologies is Compute Express Link (CXL), a new memory fabric technology that has the potential to transform the way we compute.

How? CXL is a high-speed interconnect that enables the connection of devices like CPUs, GPUs, and FPGAs to memory and storage. It is a consortium-driven standard, spearheaded by companies such as Intel, Alibaba, Google, and Microsoft, and it is designed to address the growing demand for high-performance computing and data-intensive workloads.

CXL is a natural progression of previous technologies, such as PCI Express and NVMe, which have enabled high-speed data transfer between components within a server or workstation. However, CXL takes this a step further by enabling the disaggregation of memory from compute.

The Key Use Cases of CXL

The main use case for CXL is the disaggregation of memory from compute. Traditionally, memory has been tightly coupled with the CPU or GPU, which has limited the ability to scale memory independently of compute. With CXL, memory can be disaggregated from compute, and pooled and shared between multiple computing nodes. This means that memory can be scaled independently of compute, enabling more efficient use of resources and improved performance. This can accelerate important AI/ML workloads, including the large language models.

Another use case for CXL is its ability to provide an ultra-low latency transport between computing nodes. CXL's high-speed interconnect can be used to connect multiple servers or workstations, enabling them to communicate with each other with very low latency, and shared memory. This is particularly important for financial trading workloads and real-time databases, where low latency is critical for achieving high performance.

CXL can have a significant impact on computing by blurring the boundaries between servers and enabling multiple servers to share a single pool of memory. This enables the creation of new memory-centric architectures, which can improve the performance of computing applications. Because CXL can be used to connect memory-intensive applications, such as AI and machine learning workloads, to high-speed memory pools, it can enable faster training times and more efficient use of resources.

Walking in Fibre Channel’s Footsteps

CXL is analogous to Fibre Channel, the first storage networking technology that disaggregated storage from compute in the early 1990s and created a new industry of networked storage systems and software. Fibre Channel enabled the creation of shared storage pools, which enabled more efficient use of resources and improved performance. A new industry of storage systems and software emerged because of it, and leading companies such as EMC, NetApp, Veritas, and Pure Storage were born.

Similarly, CXL enables the disaggregation of memory from compute, which enables the creation of shared memory pools. This can dramatically improve the performance of data-intensive workloads and enable more efficient use of memory resources. CXL, too, has the potential to create a new industry of memory systems and software, which will transform the way we process data.

Steve Scargall is a Senior Product Manager and Software Architect at MemVerge.

Related articles:

· How DPUs, IPUs, and CXL Can Improve Data Center Power Efficiency

· IPU Chip Offloads Networking and Some Security Tasks from CPUs

· Reducing Network Overhead with Compute Express Link (CXL)