AMD Unveils Latest Patent for Enhanced Memory Performance with Multi-Chip DRAM Approach, Significantly Benefiting APUs

AMD has recently unveiled an innovative patent focused on enhancing DRAM performance, which remarkably doubles memory bandwidth without necessitating faster DRAM silicon. Instead, this advancement leverages modifications in on-module logic.

AMD’s Patent Achieves Dramatic Memory Bandwidth Increase Without DRAM Silicon Advancements

Traditionally, upgrades to hardware involve inherent challenges, often requiring substantial architectural modifications or the reworking of logic and semiconductor technologies. However, AMD’s latest patent introduces a groundbreaking concept termed ‘high-bandwidth DIMM’ (HB-DIMM), which successfully doubles DDR5 memory bandwidth output by implementing simpler, yet effective changes. Rather than concentrating solely on the advancement of DRAM processes, AMD has ingeniously combined register/clock drivers (RCD) with data-buffer chips, enabling a significant uplift in memory bandwidth.

Diagram of a circuit labeled FIG.3 showcasing components like PLL, PC0, and PC1 with varied data paths and speeds — Diagram of AMD’s HB-DIMM Patent

Diving into the technical specifics, the patent outlines how the HB-DIMM methodology sidesteps direct DRAM enhancements. Through techniques such as re-timing and multiplexing, memory bandwidth is boosted from 6.4 Gb/s per pin to an impressive 12.8 Gb/s per pin. This doubling of bandwidth occurs as AMD utilizes the onboard data buffers to merge two conventional-speed DRAM streams into one accelerated stream directed at the processor, thus enhancing overall performance.

This technology is primarily tailored for artificial intelligence (AI) and similar bandwidth-intensive applications. Additionally, the patent suggests an intriguing implementation for Accelerated Processing Units (APUs) and integrated Graphics Processing Units (iGPUs).It proposes the use of dual ‘memory plugs’: one utilizing standard DDR5 PHY and another designated as an HB-DIMM PHY. This hybrid approach optimizes memory utilization; the larger memory pool derives from the standard DDR5, while the HB-DIMM channel is engineered for high-speed data transfer.

Diagram of a GPU architecture indicating components like command processors, memory controller, caches, and DDR memory — HB-DIMM Integration in APUs

In the context of APUs, this innovative strategy offers superior responsiveness for on-device AI tasks requiring substantial data throughput. As edge AI continues to gain prominence in conventional computing environments, AMD stands to benefit significantly from this development. However, it is worth noting that increased memory bandwidth might lead to heightened power consumption, necessitating efficient cooling solutions to manage this demand.

Overall, the HB-DIMM approach showcases great potential, effectively doubling memory bandwidth while sidestepping the need for advancements in DRAM silicon technology.

Source & Images