RDRAM Performance

[flagreen]

I saw this article linked to in the Ars forum. It compares overall system improvement with higher speed RDRAM over PC800. Impressive improvement in Sysmark scores with PC1200 RDRAM. Looks like it will really help Intel in the cpu wars.


http://www.hardwareanalysis.com/content/article/1444.6/

 

[time]

1. Their Sysmark conclusions are specious:

The SYSmark 2001 benchmark clearly shows that the gap between PC-800 and PC-1066 is where the most performance is gained. Going from PC-1066 to PC-1200 and a 150MHz FSB doesn’t scale similar to going from PC- 800 to PC-1066.

Simple arithmetic lesson. 1066/800 = +33%, 1200/1066 = +12.6%.
The performance increase was 9.7% and 6.9% respectively.

2. IMHO, they used an inappropriate graph for memory throughput. The vast majority of ops are reads, so stacking them on top of writes is misleading. They can't round either - that 9.7 should be 9.8.

3. I guess I'm out of touch here, but how did they adjust the multiplier on a P4, engineering sample or not?

4. If their observations are true, the P4 is even more bandwidth hungry than we thought. Rather than a positive development for Intel, I would say they will be in big trouble by the time they hit 3GHz.

 

[flagreen]

I'm in no position to dispute your numbers (nor do I have reason to) however I do disagree with your overall conclusion that this is not helpful to Intel. Benchmarks sell. It's as simple as that. It's not a smart way to buy, but it's true. Time after time (no pun intended) we have all seen comparison benchmarks between P4s and XPs where Sysmark numbers are used to show the P4 as being inferior to the XP. If with PC1200 these Intel marks now surpass AMD then it is going to boost Intel's sales. And sales is what it is all about.

I will take minor issue with the following;

Simple arithmetic lesson. 1066/800 = +33%, 1200/1066 = +12.6%.
The performance increase was 9.7% and 6.9% respectively.

You are assuming that overall performance scales on a 1 to 1 ratio with FSB speed increases. However this has not been established. Any system has many factors upon which overall performance in a benchmark such as Sysmark depends. Only one of these is memory performance. And it is not unusual for there to be diminishing returns on performance improvement as bus speed increases.

BTW Intel P4 ES cpus are not multiplier locked.

 

[time]

You are assuming that overall performance scales on a 1 to 1 ratio with FSB speed increases.

Going from PC-1066 to PC-1200 and a 150MHz FSB doesn’t scale similar to going from PC- 800 to PC-1066.

In fact, their figures show performance scaled better between 1066 and 1200 rather than 800 and 1066. Which is surprising to say the least. :roll:

I don't mean to sound too querulous, but I didn't realize P4 motherboards provided a way to change the multiplier. Obviously some must, and the unlocked chip must have some way of being reprogrammed accordingly? I really don't know, I'm just genuinely curious.

 

[flagreen]

In fact, their figures show performance scaled better between 1066 and 1200 rather than 800 and 1066. Which is surprising to say the least.

Huh? This doesn't appear to be the case. From what I see they are saying exactly the opposite. What have I missed? I assume we are talking about the Sysmark scores?

I don't mean to sound too querulous, but I didn't realize P4 motherboards provided a way to change the multiplier. Obviously some must, and the unlocked chip must have some way of being reprogrammed accordingly? I really don't know, I'm just genuinely curious.

I thought it was common knowledge that none of Intel's engineering sample cpu's are multiplier locked. At any rate virtually all P4 boards have the means to adjust the multiplier through the BIOS. However the BIOS selection will be overridden by the CPU if it is locked. [

 

[time]

Okay, let me express the results by showing the performance gain as a percentage of the theoretical maximum. The theoretical maximum improvement between 800 and 1033 is 29%, and between 1033 and 1200 is 16%.

Their quoted SysMark figures are 184, 202, and 216 for 800, 1033 and 1200 respectively. So the actual improvement between 800 and 1033 is 9.8%, and between 1033 and 1200 is 6.9%.

Therefore the percentage of possible gain achieved between 800 and 1033 is 9.8 / 29 = 34%, and between 1033 and 1200 is 6.9 / 16 = 43%.

So the gain appears to be growing rather than shrinking. This can make sense with cam effects in a car engine, but not here. And my point was that they claimed the opposite was happening.

Thanks for the info on P4 multipliers. I just didn't understand why motherboard manufacturers bother to include a multiplier setting, let alone in the BIOS, when only engineering samples are unlocked. Perhaps high end P4s are also unlocked? I think Intel used to do this.

 

[flagreen]

Hmm... If you analyse it as follows it dimishes as expected;

No. of mhz required per score point with PC 800 = 4.347mhz

No. of mhz required per score point with PC 1066 = 5.277mhz

No. of mhz required per score point with PC 1200 = 5.55mhz

Differential in mhz required at 1066 vs. 800 = .93mhz (21.3% over PC 800)

Differential in mhz required at 1200 vs. 1066 = .273mhz (5.2% over PC 1066)

This then indeed shows a diminishing return as expected. Or maybe it just shows that I can lie like a crooked accountant!