> Which drives obsolense of memory, price or performance is a very good argument.
> I would have to believe price is more important than performance in this case since
> no one is seeing more than 10-15% improvements in performance with DDR.
I might challenge that assessment. Going from PC133 to PC1600 -- a speed grade which has better bandwidth but worse latency than its predecessor -- performance in SysMark and Quake increased by 8 and 10 percent. We have yet to see scores of a PC2100-based system with a processor and chipset that supports the full 2.13GB/s memory (just like, for example, we have not yet offically seen benchmarks of a system with DRDRAM that can handle its full 1.60GB/s bandwidth).
I'm not asserting that we'll see these 10-15% performance increases with PC2100 memory over PC133 memory, mind you. I just wanted to note that we've currently been limited to seeing the following combinations:
A) Athlon with PC1600 because there is no 266/133MHz chipset compatible version of the Athlon and the AMD760 is synchronous.
B) Pentium III with PC2100 memory such that the chipset is limited to the bandwidth of PC133 memory.
Once again, this is not a claim or a prediction that DDR SDRAM performance will be great. I'm merely saying that we haven't seen the PC2100 SDRAM with 2.13GB/s chipset example which will be (fairly shortly) announced by AMD (and I'm making no promises here towards availability -- it's recently become standard operating procedure among almost all tech companies to make availability come several months after announcement; AMD's been pretty good in spite of this in the current year, but I presume nothing here).
Additionally ... if PC2100 SDRAM offered a 15% performance boost over PC133 SDRAM: Well, it would be fairly spectacular. To give you an example, a speed grade boost from 0.90GHz to 1.00GHz offers about a 5% performance (according, for example, to AnandTech's recent 1.20GHz Athlon benchmarkfest). That's 5% for an 11% boost in cpu speed. A 15% across the board improvement (and, once again, this is a theoretical discussion and I am in no way actually positing that this is the advantage for PC2100) for general office apps and games and soforth might be equivalent to a 35% boost in cpu frequency, or more since performance scaling is sublinear (er, that's not exactly the word I'm looking for, but you know what I mean). This would mean that, under this totally hypothetical scenario, an Athlon-1.00 with PC2100 memory could outperform an Athlon-1.30 with PC133 memory and might even come into the ballpark of an Athlon-1.40 with the slower memory configuration.
15% performance improvements for the new memory type would be awesome. 10% improvements would be damn good. 5% would be nice, but nothing to cry about and would need to be similar in price to the memory it is replacing.
BTW, a techish question for those applicable: HWP (hardware prefetching) appears to help latency, possibly drastically, at the expense of bandwidth. It is my speculation that the i840's supposed "prefetch cache" was a large factor in the specfp boost over i820, since the prefetch could utilize parts of the extra 1.6GB/s bandwidth to decrease the effective latency in the memory subsystem. Continuing on this speculation, I think I'm asserting that HWP would be a little bit more of a benefit on systems with DRDRAM than on systems with SDRAM, as the former is said to have a higher latency and, in the case of the i850, a higher peak bandwidth.
Does anybody have interesting comments about this?
-JC |
