Engineer News Network The ultimate online news and information resource for today’s engineer
The demand for AI servers is rising quickly, with these systems expected to account for approximately 65% of the server market by 2024, according to Trendforce (2024). This growth has created an urgent need for greater memory bandwidth and capacity, as AI servers now require at least 1.2TB of memory to operate effectively.
Traditional DDR memory solutions are increasingly struggling to meet these demands, especially as the number of CPU cores continues to multiply, leading to challenges such as underutilised CPU resources and increasing latency between different protocols.
Innodisk´s CXL memory module addresses these challenges by overcoming the limitations of conventional DIMM channels and significantly boosting server system performance. The module supports 32GB/s of bandwidth and supports data transfer speeds up to 32GT/s via the PCIe Gen5 x8 interface, ensuring the rapid processing capabilities essential for AI workloads. If equipped with four pieces of 64GB CXL memory modules, the server configured with eight pieces of 128GB DRAM modules can increase its memory capacity by 30% and bandwidth by 40%, which can meet the demanding memory requirements of AI servers without requiring additional DIMM slots, thereby optimising hardware architecture and reducing system complexity.
Furthermore, the CXL memory module enables memory pooling, which optimises memory resource sharing between CPUs and components, significantly reducing redundant memory usage and enhancing overall system efficiency.
The CXL memory module from Innodisk features the E3.S 2T form factor based on the EDSFF standard. This design allows for flexible memory expansion and easy module swapping within servers, ensuring seamless integration with minimal cost and complexity. CXL is an open standard promoted by major industry players, and it is expected to rapidly develop into a comprehensive ecosystem, which is critical for applications in cloud data centres, networking communications, and edge servers.
