RAM to avoid: hot, expensive and slow

TalkBack 1 of 17:: Next »

Thread View
Flat View

JEDEC: FB-DIMM was proposed by Intel back in 2003 (March, to be specific) to the JC-42 memory committee. The reason for it, as they explained in their slideware, was that server performance had more to do with total amount of memory than with either processor or memory speed.

Trouble was, as speed increases the number of DRAMs you can hang on a controller decreases. Intel's marketing projection was that in about 2005-2006 net server performance would start to decrease with each generation, causing the End Of Civilization As We Know It [1]. They also projected a net system cost adder of less than $20/module for the controller and wanted every semi company on Earth to make controller chips.

The next several meetings didn't make a lot of progress because Intel and IBM were pushing to get buy-in for a spec that they had under NDA, but wouldn't show at meetings. Finally they wraps came off far enough for critical review, and then working groups got started.

At about the same time that the JEDEC FB-DIMM spec issued, AMD was starting to seriously eat into Intel's server share and then Intel held their breath and deep-sixed the NetBurst architecture in favor of multi-core. Combined with a multi-issue controller that bought them enough performance that they didn't need FB-DIMM for a while longer, but it remained on the roadmap and will until next year or 2009, when they finally ditch the FSB.

As a strategic commitment, FB-DIMM is dead after that, since both Intel and AMD will have on-processor memory controllers.

[1] The AMD guys in the room were doing their best not to laugh out loud; they already had a solution and Intel was doing its best to badmouth it. Fast forward: next year Intel will innovate the same solution that AMD had then.; Posted by: Yagotta B. Kidding Posted on: 05/21/07 You are currently: a Guest | Members login | Terms of Use