Compute Express Link (CXL), the technology that connects memory, was one of the themes at the Future of Memory and Storage Summit held in Santa Clara last week.
CXL is an open standard for high-speed connections between processors and devices, and between processors and memory. Designed for high-performance datacenter servers, it allows processor modules to share memory.
It provides an interface based on the CXL.io protocol that uses the PCIe 5.0 physical layer and electrical interface, providing a low-latency interconnect path for memory access, communication between the host processor and devices that need to share memory resources. In other words, CXL uses the PCI express (PCIe) expansion bus, allowing you to increase memory on a PC server even when all memory slots (Dimm sockets) are full.
According to Samsung, CXL enables seamless memory sharing and memory expansion without the use of additional technical infrastructure.
CXL Memory Modules (CMMs) are intended to address the rapidly growing demand for high memory capacity in AI applications, one use case being to support the memory architecture required for large language AI models.
In a community blog post published in March, Intel noted that training large language models (LLMs) such as GPT-4, Llama 2, and PaLM 2 requires large amounts of memory and compute power.
The chipmaker explained that CXL allows server makers to take advantage of cheaper memory to reduce hardware costs. Intel said in a blog that Micron is offering a 128GB CXL module that uses DDR4 memory, the previous generation of double data rate synchronous dynamic random access memory chips.
CXL for Server Applications
Since 2022, Samsung has partnered with Red Hat to focus on developing and validating open source software for existing and emerging memory and storage products, including NVMe SSDs, CXL memory, compute memory/storage, and fabric.
At the Red Hat Summit in Denver in May, Samsung demonstrated its CMM-D1 memory model built into Red Hat Enterprise Linux 9.3, which the company says will improve the performance of deep learning recommendation models (DLRM).
This marks the first demonstration of a CXL infrastructure certified by Red Hat for Red Hat Enterprise Linux.
The demo used Samsung's open source Scalable Memory Development Kit (SMDK) memory interleaving software technology to improve memory access performance. According to Samsung, SMDK-equipped CMM-D memory with CXL support allows developers to build high-performance AI models without heavy investment. According to Samsung, the technology accelerates the speed of data processing, AI learning and inference.
Commenting on the demo, Kyeong Sang Kim, general manager of Red Hat Korea, said, “Samsung's hardware optimization for Red Hat's software underscores the value of open source technology as a critical enabler in scaling next-generation memory solutions such as CMM-D.”
In its Red Hat Ecosystem catalog listing, Samsung explains how RHEL's feature called memory tiering supports CMM-D memory modules: According to Samsung, when seldom accessed data is stored in local memory, memory tiering reduces the performance of frequently used, or “hot,” memory. RHEL's memory tiering feature is designed to allocate hot memory to the local memory tier and less frequently used memory to the CXL tier, allowing data to be migrated between tiers as needed.
