Product: Kingston ValueRAM 4 x 1 GB DDR2-800 FB-DIMM kit
website: www.kingston.com

Those infamous FB-DIMMs have finally made it into high-end desktop space through Intel's D5400XS "Skulltrail" mainboard. They got the capacity, up to 8 GB per DIMM right now - even though Skulltrail would only support up to 4 GB ones for a total 16 GB RAM capacity: enough even for Vista, I would say.

However, even on the FBD-800 variety, the overall latency - a combination of AMB memory buffer chip on the DIMM, and the memory chip latency, still affects the performance somewhat. Most FB-DIMMs around only reach 5-5-5-12 latency in usual operation. The Microns I originally got with the Skulltrail test unit could do 5-4-4-9 in the "alpha" early BIOS rev; it was slowed down to 5-4-4-12 in the "beta" BIOS, and, unfortunately, stuck with standard 5-5-5-12 in the production BIOS 0926, still the newest one on the block.

At the same latency, due to 1.8 v power and seemingly faster AMB chip, those Microns were beating the cooler 1.5 v Nanya DIMMs in Sandra bandwidth results at the same latency settings by nearly 10%. However, surprise surprise, with the new BIOS, Nanya could still do 5-4-4-9, and gain back a bit of the performance compared to the Micron. More from that review here .

Here I got Kingston's FBD-800 ValueRAM - aka cheap - parts, the lowest-capacity 4 x 1 GB kit configuration. This is not a recommended Skulltrail capacity setup, though: whoever buys this quad-channel, 8-core monster board will definitely be running a 64-bit OS (Windows XP 64 or Vista64 or Linux or Solaris) and 8 GB RAM, i.e. four identical 2 GB DIMMs, one per channel, are kind of minimum.

Since these DIMMs use lower-capacity 512 Mbit dies instead of larger & slower 1 Gbit dies on 2 GB modules, I was curious to see if we could tune down the latency some more, all that without increasing the voltage from 1.8 volts DRAM and 1.5 volt AMB chip (it is the AMB that creates most of the FB-DIMM heat, so its voltage lowered helps a lot).

And, there you go - after a few tries, I got 4-4-3-9 running fine at those default voltages. And no, it wouldn't budge even a notch better above these timings, even after I upped the DRAM voltage to 2.2 volts. And yes, this BIOS rev still forces Trasmin setting to 12 even though the chips can go 9 or even less. Here is the Sandra run now:

Compare it to the previous Micron and Nanya results - 8885/8385 MB/s and 8388/8388 MB/s respectively, with both having 93 ns latency and, umm, the Kingston units are almost matching the Micron numbers, yet requiring lower latency settings for that. The answer to the plight? Again, the AMB chip choice... it does seem to play a pivotal role in the real obtainable bandwidth of a FB-DIMM.

In summary, this is a good beginning: Kingston is expected to announce its even faster, heat pipe cooled HyperX FB-DIMMs anytime now, with OCZ and Corsair having their offerings in this class soon as well. For all of them, this quickie test has a message: the choice of AMB chips on those modules will affect their performance just as much as the memory dies and cooling used. It will be interesting to compare those results, hmm...