Intel DDR2 FB-DIMM Performance

Intel was first to adopt DDR2 memory when they had launched their LGA-775 socket nearly two years ago with the Grantsdale and Alderwood Chipsets. Intel Corporation is first again to introduce the latest in memory technology: FB-DIMM. FB-DIMM is short for Fully Buffered Dual Inline Memory Module, and is primarily designed for mission-critical server environments that require maximum performance with minimal errors. FB-DIMMs are designed to bring the best traits from DDR2 memory while combining a new point-to-point serial memory interface. Some of the key benefits for Fully Buffered DIMMs include enhanced reliability, greater bandwidth, improved scalability, and higher capacity per memory channel.

The serial memory interface for FB-DIMMs allows a greater quantity of memory modules to be used, where a traditional system would fail. The weakness of conventional DIMMs is the dedicated connection with the memory controller, and the overloading of these connections with a high memory concentration. One of the notable factors for FB-DIMMs is the adoption of an AMB (Advanced Memory Buffer) chip. The AMB controls the transmission of memory data into packets. This process is similar to that of PCI Express, SATA, and other serial technologies. The AMB chip does communicate with the next succeeding chip through a bi-directional serial interface. Meanwhile, the first AMB chip in the channel is what communicates with the memory controller. The Advanced Memory Buffer should also make it an ease for these mission-critical servers to adapt new memory standards, such as DDR3, while using the same 240-pin interface and making no other upgrades.

Another benefit of FB-DIMMs is the lower pin count for the memory controller and easier routing of multiple channels across the motherboard due to the serial bus. The serial data path of this memory should provide faster performance over current DIMMs. FB-DIMMs are also immune from increasing latencies when the memory density increases. FB-DIMMs also utilize ECC (Error Checking and Correction), and incorporated into the modules is CRC (Cyclic Redundancy Check). On top of Cyclic Redundancy Check is also a bit-lane fail-over correction feature, which aims at locating and removing bad data paths.

Current limitations of this technology are each FB-DIMM channel can only contain up to eight modules, and only six memory channels are allowed per memory controller. Even so, that is a possible 48 FB-DIMMs per memory controller. Now some food for thought, with 48 FB-DIMMs each offering up 16GB of memory (which will soon be available), the total capacity would be 768GB. The governing body over semiconductor engineering standards, JEDEC, had only finalized the specifications for FB-DIMM last month.

Even with the memory being buffered and the Cyclic Redundancy Checking, FB-DIMMs are theoretically faster than traditional server memory, which is the focus of the entire article. Of the speed advantages include the bi-directional serial interface being able to read and write data simultaneously, greater number of memory channels, and more. We will be looking at its Linux memory performance in the server environment compared against existing DDR1/DDR2 desktop and server modules, as well as looking after the effective memory bandwidth when more memory channels are enabled.

Feel free to discuss this article in our Phoronix Forums.