DDR5 12.8Gbps MRDIMM IP: Cung cấp năng lượng cho tương lai của AI, HPC và trung tâm dữ liệu

The demand for higher-performance computing is greater than ever. Cutting-edge applications in artificial intelligence (AI), big data analytics, and databases require high-speed memory systems to handle the ever-increasing volumes and complexities of data. Advancements in cloud computing and machine virtualization are stretching the limits of current capabilities. AI applications hosted in the cloud rely on fast access and reduced latency in memory systems, which is amplified by an increasing number of CPU and GPU cores.

Introducing the DDR5 Multiplexed Rank DIMM (MRDIMM), the next-generation memory module technology designed to meet the needs of high-performance computing (HPC) and AI in cloud applications. By leveraging existing DDR5 DRAM memory devices, MRDIMM modules not only double the DRAM data rate but also maintain the RAS capabilities of the industry-proven RDIMM modules, setting a new precedent for memory module performance.

Let’s compare RDIMM and MRDIMM modules using the same DRAM parts. Today, high-speed production DDR5 RDIMM modules run at 5600Mbps. Those modules use DDR5 DRAM parts, which also run at 5600Mbps. An MRDIMM module using the same DDR5 5600Mbps DRAM parts will run at a blazing 11.2Gbps.

One key metric for best-in-class performance, low bit error rate (BER), and ease of adoption is the eye diagram. The eye diagram illustrates at-speed system margin and accurately represents DDR system quality when captured with a pseudo-random binary sequence (PRBS)-like pattern. The diagram below illustrates Cadence’s 3nm silicon write eye diagram for DDR5 MRDIMM IP running at 12.8Gbps.

Cadence 3nm DDR5 MRDIMM 12.8Gbps test chip write eye diagram, design kit is available today.

The eye diagram is captured using a PRBS-like pattern, incorporating a package and system board representative of a typical MRDIMM channel. Using PRBS-like patterns is crucial for capturing accurate eye diagrams. Repetitive clock-like data patterns create deceptively “open eyes” that do not reflect the real system performance. Effects like intersymbol interference, simultaneous switching, reflections, and crosstalk are not accurately reflected in the eye diagrams for parallel interfaces like DDR using non-random data streams. Relying on improperly captured eye diagrams inevitably leads to a significantly worse real system BER than conveyed by that eye diagram.

Doubling the DDR5 RDIMM data rate is challenging. Achieving high performance while optimizing for area and power requires multiple design techniques. Feed-forward equalization (FFE), decision feedback equalization (DFE), continuous-time linear equalization (CTLE), and T-coils are required to reach 12.8Gbps MRDIMM data rates in multi-channel systems. Building a production-worthy 12.8Gbps DDR5 MRDIMM IP requires engineering expertise that comes from many generations of memory interface design and production experience. Cadence has developed this expertise through multiple DDR5/4, LPDDR5X/5, and GDDR6 designs in different technology nodes and foundries. For instance, Cadence’s GDDR6 IP is available in three foundries and ten process nodes, with mass production at speeds exceeding 22Gbps.

For your next project, consider DDR5 12.8Gbps MRDIMM, a technology that not only doubles the bandwidth of DDR5 RDIMM but also promises rapid proliferation into next-generation AI, data center, HPC, and enterprise applications. With its cutting-edge capabilities, the Cadence DDR5 12.8Gbps MRDIMM IP is ready to power the future of computing.