RAM to avoid: hot, expensive and slow

Has Intel ever met a technology they couldn't make hotter, costlier and slower? The latest victim: DRAM.

Has Intel ever met a technology they couldn't make hotter, costlier and slower? The latest victim: DRAM. Here's what you need to know.

Intel is pushing something called Fully Buffered DIMMs (FB-DIMM). You see them on some servers and high-end PCs like the Dell 690 and the Apple Mac Pro that use the Intel 5000 series chipset. The best web price is $82/GB versus $32.50 for PC2-5300 DDR. Ouch! And each FB-DIMM consumes an extra 3-5 watts for the buffer/control chip. More ouch.

What Intel claims on their website

-Provides over 3 times higher memory throughput† allowing for superior application responsiveness -Enables increased capacity and speed to balance capabilities of dual core processors

† Based on the Dual-Core Intel® Xeon® Processor 5000 sequence (Dempsey)/Blackford chipset with four memory channels each running at 667MHz results in a theoretical throughput of 21.3 GB/s as compared to an Intel® Xeon® Processor with 800MHz FSB with Intel® E7320 chipset with 2 memory channels at 400MHz gives 6.4GB/s theoretical throughput.

Hold tight to your wallet when you see "theoretical" in a tech spec.

Well, they did achieve higher capacity Anand Lal Shimpi at AnandTech reports on what FB-DIMMs actually deliver in a review of a Mac Pro, which uses an Intel-designed motherboard.

. . . it actually takes longer to access main memory than the Core Duo processor in the MacBook Pro. This is much worse . . . [since] the MacBook Pro features a 667MHz FSB compared to the 1333MHz FSB (per chip) used in the Mac Pro. . . .

It's not Apple's fault, but FB-DIMMs absolutely kill memory latency; even running in quad channel mode, the FB-DIMM equipped Mac Pro takes 45% more time to access memory than our DDR2 equipped test bed at the same memory frequency. Things don't get any prettier when we look at memory bandwidth either.

. . . With four FBD channels, the best we're able to see out of the Mac Pro is 4.292GB/s, compared to the 6.782GB/s of bandwidth our dual channel Core 2 testbed is able to provide. . . .

FB-DIMMs are simply not good for memory performance. . . .

So what, exactly, is the upside? To be fair, FB-DIMMs have a few advantages.

  • Capacity. They may be slow, and get even slower as you increase the number of DIMMs, but you can build a system with 32 or even 64 GB of capacity. You just can't afford it.
  • Cheaper printed circuit boards. FB-DIMMs use a serial interconnect, so there are fewer wires on a circuit board, which makes them cheaper to build. Me, I'd rather spend $50 more for the PC board and $50 less for every DIMM.
  • Reliability. FB-DIMMs use CRC and other techniques to reduce potential problems. You know all those problems you have with memory errors? Me neither.
  • Simultaneous read/writes. This ameliorates some performance problems if you have a workload with lots of overlapped reads and writes.

The Storage Bits take Intel calls FB-DIMMs ". . . a long-term strategic direction for servers." "Strategic" is marketing-speak for "makes no sense today." AMD has passed on them for Opterons, and the cost and power issues are in the process of killing FB-DIMMs even for servers.

Intel just started manufacturing mobo's for Google and you can be darn sure FB-DIMMs aren't on those either. 4 GB DDR2 DIMMs are just now making it to market and while they are slightly more expensive than even FB-DIMMs today, I'd expect prices to drop pretty fast as production ramps. In short, the strategic direction for FB-DIMMs is oblivion.

The more interesting question is why Intel keeps making poor architecture decisions such Itanium, NetBurst and RDRAM and now FB-DIMMs. I think most of these designs have come out of Intel's Hillsboro, OR plant, so I suspect the problem resides there as well. Perhaps Intel's former supercomputer group diffused a taste for grandiose and uneconomic architectures throughout the engineering team.

Here's a tip: don't construct expensive workarounds for problems that are four years out. You aren't that good. AMD has figured out a way around this, and you could have too, if you'd given it a chance.

Comments welcome, as always.