Dual DDR is coming, and the RDRAM advocates don’t like it.
Understand The Real Matchup
Some RDRAM geektards
have taken some Granite Bay benchmarks and are jumping up and down saying that dual DDR is a big disappointment and this proves RDRAM is better.
This article will explain why this arguments has some big holes in it and tell you what you need to look for
to make an intelligent decision about the products.
DDR Runs Faster
There’s something noteworthy about these RDRAM/DDR comparisons. They’re always run at 133MHz.
What this does is strips away DDR’s main advantage over RDRAM. It just can run at higher speeds than RDRAM.
You would be a fool to choose dual DDR over RDRAM if all you were going to do is run it at 133MHz.
No, you’re going to run it at 180 or 200 or 220, or whatever you can get out of it. Whatever “whatever” is, it’s going to be a lot faster than the 150ish speed you can get out of RDRAM.
That’s the real-life comparison: not 133 versus 133.
Saying that 133 versus 133 is the only “fair” test is like a Mac user telling you to slow your processor down to 1.4GHz to test against his latest and greatest G4 processor.
Just like the RDRAM person, he’ll say that this “fair” test will prove that the Mac G4 processor is more efficient cycle for cycle. In both cases, this is completely true and completely irrelevant in real life. Both the latest x86 processors and DDR run much faster in real life than in the test.
The first dual DDR PIV board a lot of overclockers will buy will be the Canterwood motherboards. There, the top default speed is 200MHz. You would be crazy to buy RAM for it and get anything less than PC3200 for it.
That’s the point of real comparison for this audience.
DDR Write Performance
Before people send me this link, let me do it to myself: http://www.aceshardware.com/read.jsp?id=50000346.
This page shows that while dual DDR does a little bit better than RDRAM in read operations, it does a lot worse in write and mixed read and write operations.
I have no reason to doubt the reviewer’s measurements, but this is another one of those “completely true and completely irrelevant” numbers when you look at a real world situation.
Run that test at 200MHz, and the results will almost certainly turn into dual DDR doing a lot better than RDRAM in reads and a little worse in writes.
How Much Does Memory Matter?
If you look at the benchmarks in that article and compare performance between “fast” DDR 266 and 32-bit RDRAM, the performance difference in the games and real applications is: 2% or less.
Yes, a few of the SPEC ViewPerf tests shows greater differences, but those particular tests have a history of greatly exaggerating any differences between two platforms.
So a 60% difference in write speed hardly shows up in the real-world.
The point is not to make excuses for DDR; the point is to demonstrate that memory differences don’t matter much in the real world, any which way.
If the FSB were kept at 133MHz, and the memory were somehow run at 200MHz at fast settings, I’d bet the dual DDR would look great in the memory benchmarks, but in the real world, instead of being a tiny bit slower than RDRAM, it would be a tiny bit faster instead.
So these memory articles, including this one, are much ado about little.
One Unanswered Question
What we don’t know yet about dual DDR systems is how much improvement will occur when you run the FSB at very high speeds.
In short, given the same MHz, does running FSB and memory at 200MHz do better than, say, running FSB at 150MHz and memory at 200MHz?
When you increase the FSB, you expand the system bandwidth, which means that fast memory gets a chance to do some good. In Athlon systems, you have to increase the FSB to let faster memory do its job.
However, the PIV is a different beast.
In a single DDR channel PIV system, you can run memory faster than the FSB and have it do you good because the bandwidth from a single stick of DDR fits quite easily into the PIV’s overall bandwidth.
Dual DDR is a different story. In theory, running a dual DDR system at 150MHz while running memory at 200MHz could create a bottleneck because memory bandwith approaches overall FSB bandwidth.
On the other hand, it may well not matter in real-life for a few possible reasons, most notably CPUs running at X speed just can’t take advantage of extra bandwidth.
That’s the result you need to see before you decide to buy a Canterwood-based system. If running at 200MHz FSB doesn’t do anything for you, your best overclocking chance will likely be with a 133MHz PIV.
If running at 200MHz does make a considerable difference in performance, than the nod will go to the new 200MHz PIV, and overclocking them will likely be a considerably more difficult task.