You are here

FBDIMM Not Being Extended to DDR3

The Intel Developers Forum is naturally focused on processor related topics, but the close tie between the processor performance and the memory configuration means that there are always some important memory topics that are presented. The memory topic of highest interest this year was the resolution of fully buffered DIMM for DDR3.

Power consumption related to both the processing function as well as for the cooling of the semiconductor components are both becoming more important. In the overall scheme of a processor’s architecture, the way in which the processor uses the various memory elements as well as the coordination between the memory and the processor is critical to the overall cost and performance of the system design. The system-level design objectives impact many of those architectural considerations.

One of Intel’s design objectives relative to DRAM memory features was the ability to extend the capacity of the memory for high-performance applications with as little disruption to the architecture as possible. In order to support that objective, Intel supported Fully Buffered DIMM (FBDIMM) and the Advanced Memory Buffer (AMB) that was required on each of the modules.

However one of the performance tradeoffs was associated with the buildup of the heat from the AMB. Recent studies suggested that the heat buildup resulted from the fact that the AMB remained fully powered and functional regardless of the data processing load, and that the total power consumption of the system therefore did not seem to scale linearly based on the processing load.

The question that remained unanswered was whether Intel would continue to support the AMB configuration for DDR3 or whether one of the other well-established module configurations would be chosen.

While the performance attributes of easily extending the capacity of the memory configuration are still desirable characteristics, Intel gave a clear impression at IDF that FBDIMM would not be extended to DDR3. A quick survey of the AMB manufacturers showed that most had already accepted this change in Intel’s DDR3 strategy, but that adequate support for DDR2 FBDIMM would be available.

Semico Spin

Regardless of the performance advantages for fully loaded systems, the cooling of the FBDIMM has been a challenge. Most architectural considerations are a compromise between the array of possibilities, and the decision not to extend FBDIMM to DDR3 is one more example of the center of gravity shifting away from the highest performance systems and toward a broader user base. The different memory requirements for ultra-mobile, mobile/laptop, desktop, and high-end applications make it more difficult each year for a single memory configuration to support an ever-widening set of performance requirements.

We believe that recent efforts to establish a separate standard for mobile applications would effectively benefit the high-end applications by separating high-volume DRAM into two different tracks. A new standardization track for ultra-mobile and laptop applications could focus on tradeoffs that blend the lowest power consumption with other considerations such as density or performance. A separate standardization track focused on higher performance for larger memory arrays could then balance a different set of design criteria. Traditional market pressures tended to keep the memory manufacturers in fierce competition based on the lowest manufacturing costs for the same single product. We believe that the range of applications supported by “commodity DRAM” is too broad to continue relying on a single standard and that both OEMs and memory manufacturers are more receptive to the new dynamics associated with a wider selection of DRAM configurations.

Twitter