Today, data centers face challenges to keep up with the growing requirements of AI and HPC workloads. Increased compute and memory demand for AI applications highlights a major challenge in the market known as the “memory wall.” In short, the memory wall describes the growing gap between compute capability and memory capacity and bandwidth in the data center. Enhancements and improvements are necessary to meet the needs of memory-centric data center architecture.
CXL provides a unique fabric-attached memory capability to address some of these issues by facilitating memory pooling and sharing between multiple nodes, improving memory utilization, bandwidth, performance, and capacity for AI and HPC applications. Shared memory using CXL enables efficient, low-latency communication between processors in a data center by allowing multiple processors to access the same block of memory and eliminates the need for data sharding. Furthermore, with CXL fabric-attached devices, applications can coherently access memory, driving important TCO benefits for in-memory databases and compute-bound AI & HPC workloads.
Micron is looking forward to highlighting the benefits of CXL technology by showing a live demonstration of CXL fabric-attached memory sharing at Supercomputing 2024 (SC24).
Showcasing the Benefits of Fabric-Attached Memory Sharing
During SC24, we will demonstrate the capabilities and benefits of memory sharing across a CXL fabric with our CZ122 Memory Expansion and Fabric Attached Memory File System (FAM-FS) open-source software. Our demonstration, which includes a CXL Switch from Xconn Technologies and servers with Intel Xeon 5 Emerald Rapids CPUs, will present how a RocksDB application running on two different servers can run queries on a database simultaneously to access shared memory enabled by CXL fabric and FAM-FS. By loading a RocksDB database across 14 shared CXL modules and using the FAM-FS interleave feature, we will show how two separate servers can complete query operations in parallel on the database residing on shared memory enumerated in DAX mode, essentially doubling the performance.
Our demonstration will also highlight the key use cases for our FAM-FS, which allows applications to access a file in shared memory across all machines, helping data centers overcome the memory wall through improved TCO and memory efficiency.
View our live demonstration in the CXL Pavilion (Booth #1807) during SC24, from November 19 – 21. If you’re not attending SC24, you can view the demo in action on YouTube here. You can also visit our GitHub repository to experiment with the FAM-FS application.
Micron’s CXL Memory Expansion Modules
Micron’s CZ122 memory expansion modules use CXL to enable server OEMs to scale, integrate, and expand their memory capacity, delivering enablement for a range of application workloads. By optimizing performance beyond direct-attached memory channels and supporting fabric-attached memory applications, data centers can experience increased flexibility, memory bandwidth, and capacity as well as TCO benefits.
Micron CZ122 CXL Memory Expansion Modules
To learn more about Micron’s CXL Memory Expansion products, visit www.micron.com/CXL. You can also subscribe to our CXL technology enablement page (TEP) to access additional information about our solutions. More information about FAM-FS can be found on our Micron CXL Memory Resource Kit Repository.
If you’re interested in learning more about Micron’s production-ready CZ122 Memory Expansion device, we encourage you to contact a Micron representative by visiting our website.