Searching for the Memory Holy Grail - Part 2
by Wesley Fink on August 26, 2003 11:11 PM EST- Posted in
- Memory
Memory Configuration (continued)
Our benchmarks in Part 1 were consistent with the configuration recommendations in the Intel White Papers for Memory Configuration of the 875 and 865 chipsets. We concluded with the following performance charts based on our testing and the information provided by Intel. We differ from the Intel White Paper charts only in the first and second positions for the Intel 865 chipset. When Intel published their White Papers, we doubt that they had any notion that most of the 865 motherboards being sold would have some version of PAT (Performance Acceleration Technology) on board. Our testing of the ASUS, AOpen, and DFI PAT-enabled boards shows that the PAT 865 boards behave more like 875 boards. We have confirmed that the first and second positions on PAT-enabled 865 boards are, as we report, below — with four DS DIMMs performing faster than two DS DIMMs.
| DDR400 (1:1) Performance | DIMM Configuration | Single-Channel or Dual-Channel |
| 1 | 4 DS | Dual Channel |
| 2 | 2 DS or 4 SS | Dual Channel |
| 3 | 2 SS | Dual Channel |
| 4 | 4 SS/DS Mixed Matched Pairs |
Dual Channel |
| 5 | Any DS | Single Channel |
| 6 | Any SS | Single Channel |
| DDR333/266 Performance | DIMM Configuration | Single-Channel or Dual-Channel |
| 1 | 2 DS or 4 SS | Dual Channel |
| 2 | 2 SS | Dual Channel |
| 3 | 4 DS | Dual Channel |
| 4 | 4 SS/DS Mixed Matched Pairs |
Dual Channel |
| 5 | Any 1 or 2 DS or SS (1 DIMM or 1 in each Channel) |
Single Channel |
| 6 | Any 1 SS | Single Channel |
While SiSoft Sandra UNBuffered Memory Test results demonstrated real differences in performance among these memory configurations, many of you asked what we might see in real-world performance differences with these different memory configurations.
Some of the configurations were easier to test than others. Kingston had provided us with four Single-Bank (single-sided) DDR500 DIMMs that allowed us to look at performance differences in two SS DIMMs versus four SS DIMMs. We looked at performance of 1000FSB/DDR500 at 3-4-4-7-1 memory timings at 2.65V.
| 1000FSB (3.0GHz)/DDR500 Performance — Kingston PC4000 2 x 256 MB SS DIMMs vs. 4 x 256 MB SS DIMMs |
||||
| Number of Single-Bank DIMMs | Quake3 fps |
Sandra UNBuffered | Sandra Standard Buffered |
Super PI 2M places (time in sec) |
| 2 | 386.30 | INT 2821 FLT 2786 |
INT 5830 FLT 5748 |
109 |
| 4 | 394.57 | INT 3218 FLT 3195 |
INT 5841 FLT 5818 |
107 |
While the differences in Quake3 frame rate and Super PI times are small, keep in mind that the only change here is using four SS DIMMs instead of two SS DIMMs. As predicted by Sandra UNBuffered Memory Test, the four SS DIMMs perform better than two SS DIMMs.
Comparing two SS DIMMs to two DS DIMMs was done using 2 x 256 MB SS OCZ PC4000 and 2 x 512 MB DS PC4000 at timings of 2.5-3-4-6 at 2.65V. It is almost impossible to keep DIMM capacity the same, which would be the most accurate test, as none of the samples using the same memory chips would have the same exact capacity in SS and DS configurations.
| 1000FSB (3.0GHz)/DDR500 Performance — OCZ PC4000 2 x 256 MB SS DIMMs vs. 2 x 512 MB DS DIMMs |
||||
| Number of DIMMs & Configuration | Quake3 fps |
Sandra UNBuffered | Sandra Standard Buffered |
Super PI 2M places (time in sec) |
| 2 SS | 392.30 | INT 2918 FLT 2926 |
INT 5761 FLT 5868 |
109 |
| 2 DS | 400.10 | INT 3282 FLT 3324 |
INT 5965 FLT 5934 |
106 |
Again, the differences are small but real in Quake3 and Super PI, but with 2 DS DIMMs performing better as we had demonstrated in Part 1 with the Sandra UNBuffered Memory Test.
The fastest configuration should be 4 Double-Bank DIMMs on the Intel 875 chipset motherboard. Our first efforts to test this configuration, with 4 x 256 DS OCZ 3700 GOLD modules did not show 4 DS DIMMs any faster than 2 DS DIMMs on the ASUS P4C800-E. This was a real puzzle considering that other predicted configurations were easily verified with Quake3, Super PI and other benchmarks. When we looked deeper, the problem became clear. In every configuration that we had tested with the ASUS P4C800-E, the Performance Mode showed “enabled” but with 4 DS DIMMs at any FSB speed faster than 800, CPU-Z 1.18C shows Performance Mode “disabled”. Our trial was failing here because we were comparing 2 DS DIMMs with PAT on to 4 DS DIMMs with PAT off. We have asked ASUS for the reason behind why PAT appears to be disabled at 4 DS DIMMs above 800FSB.
Since we were mainly concerned with comparing 2 DS with 4 DS DIMMs under the same conditions, we ran 2 DS vs. 4 DS tests at 800FSB/DDR400, so that PAT was enabled in both two and four DIMM setups. We compared prototypes of high-speed PC3700 memory that we tested, looking specifically at the performance of 2 x 512 MB DS to 4 x 512 MB DS at DDR400. Timings were 2-2-4-8 at 2.75V.
| 800FSB (2.4GHz)/DDR400 Performance 2 x 512 DS vs. 4 x 512 DS |
||||||
| Number of DS DIMMs | Quake3 fps |
UT2003 Flyby fps |
UT2003 Botmatch fps |
Sandra UNBuffered | Sandra Standard Buffered |
Super PI 2M places (time in sec) |
| 2 | 321.4 | 196.30 | 68.31 | INT 2683 FLT 2722 |
INT 4704 FLT 4691 |
133 |
| 4 | 324.5 | 197.60 | 69.38 | INT 2830 FLT 2923 |
INT 4678 FLT 4717 |
132 |
While the differences here are quite small, they still show 4 DS modules performing better than 2 DS modules. We suspect that we would see larger differences at higher speed, as we have in our other comparisons.
We have confirmed with game benchmarks and a number-crunching benchmark that the best 875/865 memory configurations can also be seen in real-world benchmarks. Because memory performance is only one component of game performance and overall system bandwidth, the effect on these benchmarks is, as expected, smaller than the variation seen in benchmarks that measure only memory performance, like Sandra UNBuffered Memory Test.
Facebook
Twitter
RSS
77 Comments
View All Comments
Anonymous User - Thursday, August 28, 2003 - link
I am tired of setting the memory timing and bench mark all the time. Is there a program there which can tell me what kind of results I would get? Say if I can increase my CPU by 5 MHz but have to set back my memory timing a bit, which way should I go?oldfart - Thursday, August 28, 2003 - link
Here are some reviews comparing tight timings Vs loose:http://www.hardtecs4u.com/reviews/2003/ddr400_roun... (need language translator)
http://www.octools.com/index.cgi?caller=articles/c...
http://www.3dxtreme.org/Corsair_xms3700_twinx_p1.s...
Anonymous User - Thursday, August 28, 2003 - link
I think this was an extremely helpful and thorough review. There was one comparison, though, that I would find most helpful and haven't found anywhere. I'm currently debating the importance of running synchronously, and thus found the section "Does memory speed really matter in the real world" extremely interesting. However, I would have greatly preferred one additional test -- running 1066FSB at 3:2 and 5:4 with memory with tight timings (2-2-2-5), since my real debate is whether to buy PC3200 or PC3500 with tight timings and run at 5:4 or 3:2, or PC4000 with loose timings and run at 1:1. While I expect that the synchronous memory would result in better performance, I'd really like to know how much better, since PC4000 memory is expensive!Thanks,
Steve
Dennis Travis - Thursday, August 28, 2003 - link
Great Review Wes. Keep Em coming. I am not "PAID" to say this either. I wanted to. I am getting nothing for it either. Just the satisfaction of telling Wes I loved his review.Anonymous User - Thursday, August 28, 2003 - link
Great review Wesley. Nah I'm not paid to say this, I just enjoyed the review!Anonymous User - Thursday, August 28, 2003 - link
I have tested Kingston HyperX RAM at 1:1 3-4-4-8 @ DDR500, and 5:4 2-2-2-5 @ DDR400 at the same FSBs on a P4P800, with MAM Enabled and Turbo performance mode in both cases. While the 1:1 gets about 3-5% better Sandra bandwidth scores (buffered and unbuffered), SuperPI completes about 1.5% sooner at the 5:4 settings.So real-world performance may be slightly better at 5:4, but you won't win any Sandra bragging rights with it.
--MeowChow
Anonymous User - Thursday, August 28, 2003 - link
oldfart - Wednesday, August 27, 2003 - link
Wow, looks I'm not the only guy who wants to see250 FSB
1:1 3-4-4-8
5:4 2-2-2-6
type of testing. I've seen several reviews that show the lower latency ram @ 5:4 to be faster.
Part 3??
Anonymous User - Wednesday, August 27, 2003 - link
Hey Wesley,Thanks for all the good info...
Any chance you could test Various FSB's
5:4 2-2-2-5 vs the same FSB at 1:1 2.5-4-4-7
It would be great to show the readers how the new PC4000 REALLY compares to older slower low latency RAM, Mushkin PC3500 level2 would be perfect for that.
Now that would be a seriously good Anandtech caliber review. :D
Wesley Fink - Wednesday, August 27, 2003 - link
#32 and #40 -Mushkin did not have a product in our hands when the review was done. In fact I completed a review of Mushkin PC3500 Level II just a couple of days ago, and compared it's performance to ALL the memory in this review at DDR400. I also tested Adata DDR450, which did not meet our requirement of running at DDR500, but DID perform well at DDR400.
The reviews should be up here shortly. The Mushkin did VERY well at DDR400 to DDR450. Mushkin is also about to release DDR500 - but they did not have a product ready in time for our review. We WILL be testing it as soon as it is available if time allows.