http://www.theinquirer.net/inquirer/news/808/1017808/intel-fb-dimms-to-offer-real-memory-breakthroughsIntel FB-DIMMs to offer real memory breakthroughs
Part One Fully buffered DIMMs take shape
Monday, 5 April
2004, 06:57
THE LAST TIME Intel tried to change
the way a PC used memory, it relied on external outfits to do the
gruntwork,
and just licensed it. For a number of reasons, it didn't work out too
well, but this article is not about rehashing
history. At Spring IDF, Intel took the wraps off of what is the next
big thing, Fully Buffered DIMMs, or the slightly
easier to type FB-DIMMs.
The architecture promises to ease a lot of the problems with the current DDR and Registered DDR configurations without jacking the bill sky high. Up until now, there were solutions for capacity, speed, and cost, but rarely two, let along three in the same architecture. Did Intel achieve the proverbial memory hat trick? Did it add anything new? Well, I have not seen it running, and probably will not for a while, but until then, the IDF presentation by Pete Vogt had me pretty well convinced that it came damn close. Let's begin with the problems of current memory technologies. Today that technology is DDR. As speeds rose from DDR-200 to DDR-400, the number of DIMMs dropped precipitously. Four DIMMs, a number that is quite do-able at DDR-200 is not feasible at DDR-400. In fact it is very hard to find a DDR-400 board with more than 2 DIMM slots. Not a good start for 64-bit capable systems. Add in Registration, and you expand things a bit, you can probably get four DDR-400 Registered DIMMs on a channel, but you take a one cycle latency hit because of it. Capacity for speed, something we will get too more later on. The next way to add a slot is to change the architecture, as with the transition from DDR to DDR2, a very different architecture in a lot of ways, despite the name. While it may add capacity and the ability to ramp speeds to DDR2-800, it adds latency. Since benchmarks of DDR2 systems are few and far between, I can't say how much of a real world impact this has, but it is a potential downside. So, again, you get capacity for speed, or at least a type of speed. To make matters worse, the added capacity with DDR2-400 will drop with each speed bump. By the time things get to DDR2-800, you are probably going to be back where you started, two DIMM slots per channel. The chip density will have shot up by then, and 4GB and higher DIMMs will be fairly common, so the amount of RAM will increase enough for desktop use. On the server side, that simply won't cut it, and the lack of DIMM slots here will bring the proverbial pain. What is the problem? Basically, it is the shared bus architecture of the DIMM itself, a stub-bus configuration. What is that? Imagine a comb, with the bus being the spine of a comb, a dram on each tine, and the controller on one end. Each point where the tine touches the spine, you get an electrical discontinuity. Each of these little discontinuities is like a speed bump, it causes the signal to degrade. Just like those nasty pieces of roadway, each of these stubs causes more problems the faster the signal goes. The simple solution is to decrease the number of DRAMs on the bus, so that while cranking up the speed makes each discontinuity worse, it means there are fewer of them. Speed or capacity, pick one.
The FB-DIMMs resemble the current DDR DIMMs that are on the market, the only exception is the square chip in the center where the two chips for Registered DDR belong. You also may notice a distinct lack of pins at the bottom. Other than that, it looks like a plain old DIMM, but the magic is there. µ Part Two to follow, tomorrow |
This
is manageable for desktop uses, but for servers, a radical rethink is
necessary. Enter Intel and the FB-DIMM. The goals of this architecture
were to make a memory subsystem with vastly
higher capacity, higher performance, and adding significantly to the
reliability, all at a reasonable cost. While
FB-DIMMs may not be out yet, at least according to the presentation I
saw, the people working on it could have pulled
it off.