Channel Analysis for a 6.4 Gb s −1 DDR5 Data Buffer Receiver Front-End

• Published in: Advances in radio science 2017-01, Vol.15, p.157-161
• Main Authors: Lehmann, Stefanie, Friedel Gerfers
• Format: Article
• Language: eng ; ger
• Subjects: Architecture; Backplanes; Bandwidths; Buffers; Compensation; Computer architecture; Computer memory; Crosstalk; Data transmission; Insertion loss; Noise; Power supplies; Specifications
• ISSN: 1684-9973; 1684-9965; 1684-9973
• DOI: 10.5194/ars-15-157-2017

In this contribution, the channel characteristic of the next generation DDR5-SDRAM architecture and possible approaches to overcome channel impairments are analysed. Because modern enterprise server applications and networks demand higher memory bandwidth, throughput and capacity, the DDR5-SDRAM specification is currently under development as a follow-up of DDR4-SDRAM technology. In this specification, the data rate is doubled to DDR5-6400 per IO as compared to the former DDR4-3200 architecture, resulting in a total per DIMM data rate of up to 409.6 Gb s−1. The single-ended multi-point-to-point CPU channel architecture in DDRX technology remains the same for DDR5 systems. At the specified target data rate, insertion loss, reflections, cross-talk as well as power supply noise become more severe and have to be considered. Using the data buffer receiver front-end of a load-reduced memory module, sophisticated equalisation techniques can be applied to ensure target BER at the increased data rate. In this work, the worst case CPU back-plane channel is analysed to derive requirements for receiver-side equalisation from the channel response characteristics. First, channel impairments such as inter-symbol-interference, reflections from the multi-point channel structure, and crosstalk from neighboring lines are analysed in detail. Based on these results, different correction methods for DDR5 data buffer front-ends are discussed. An architecture with 1-tap FFE in combination with a multi-tap DFE is proposed. Simulation of the architecture using a random input data stream is used to reveal the required DFE tap filter depth to effectively eliminate the dominant ISI and reflection based error components.