I think the review sites have been wrong w/ memory timings for AMD64's

B

Barnaby

Supreme [H]ardness
Joined
Aug 7, 2001
Messages
5,280
Ok, so if anyone has followed the review sites for awhile, it's been standard to say on an AMD rig you wanted 2-2-2-10 for the best performance.. It's been like that since NF2 days, the same with the NF3, and now with the NF4 we see it go to 2-2-2-7..

So, blindly accepting this as the way it is, I set up my NF3 rig with 2-2-2-10 timings, as this should give the best performance..

One of the easiest ways you can see what your memory performance is, is to run memtest. It shows how many MB/s your L1/L2 cache does, and your ram.. Anandtech states this is typically the way they check to see which settings give the best performance..

Today I run some tests myself, just to see myself, as I needed to kill some spare time.. I used the latest version of memtest, 1.51, and checked it out..

My rig is a MSI Neo2, 3500+, OCZ 3200EL Platinum rev 2

So with 2-2-2-10 timings, memtest says I score a 2540 MB/s..
2-2-2-5 gives me 2698...

Oops.. If I'm not mistaken that means 2-2-2-5 gives better performance.. so I run through the rest of the settings.. here's a break down

2-2-2-5 = 2698
2-2-2-6 = 2618
2-2-2-7 = 2618
2-2-2-8 = 2540
2-2-2-9 = 2540
2-2-2-10 = 2540
2-2-2-11 = 2466

This goes in line with what should be happening in theory.. The lower the memory timings, the more performance you get..

To see if this is just a fluke, I run through some other benchs

EVEREST
2-2-2-5 = 5827 read/ 2233 write / 43.9 latency
2-2-2-10 = 5876 read/ 2507 write / 44.4 latency

hmm.. so everest is saying i have more memory bandwidth, but with reduced latency.. let's see what Sandra says

Sandra 2004
2-2-2-5 = 5799 int/ 5734 float
2-2-2-10 = 5785 int/ 5821 float

Hmm.. so sandra is kinda backing up Memtest alittle bit more..

So, just to get some non synthetic numbers, I try Q3 with the latest pointrelease

2-2-2-5 = 296.6
2-2-2-5 = 295.2

so, I do get slighty better performance in Q3 with 2-2-2-5.. Granted I ran the tests with 4xAA and 8xAF, but there was still a difference..


So, it would almost seem that all this time, atleast with the NF3, that 2-2-2-10 timings IS NOT the best option... I think that some benchs in their testing might be off, where it's mis-reporting the info.. Memtest is probably as raw as a program you can get, that doesn't get influenced by drivers, OS, etc.. And a realtime bench like Q3, which is VERY memory sensitive, does show a difference..

I'm going to re-run Q3 @ 1024x768 and no filtering real quick just to see if there is any difference, but so far I'm thinking lower does after all mean better for AMD64 rigs..
 
Ok, just finished running Q3 on 1024x768 with no AA or AF

2-2-2-5 = 330.54
2-2-2-10 = 327.27

=============================

So, a small difference to be sure, but it cleary is showing that 2-2-2-5 gives better performance..

I dont' even count the results of Everest or Sandra, as everytime I restart to change the timings, they give me very different scores..

But Memtest and Q3 give very consistent results..

==============================

dave_graham: that article was my reason for checking this out on my own system.. And from my own testing I have found this NOT to be true. I'm thinking Anandtech is using an older version of Memtest, as the later ones have adressed a few bugs for AMD64 chipsets..

If anyone else can try and confirm this on their own systems, I'd be intersted in seeing if this whole time, a higher tRAS settings was all based off incorrect reporting on different benchmark software..
 
yeah, i think it depends on the memory you use, the speed your ram is at, and the cas latency. there have been a few threads on this.. though none recently.

with my system, with cas2 i get the most bandwidth at a tras of 8 when at 200mhz.. haven't tested other cas or speeds yet though.. i'll try to remember to do that when i get back to school
 
well your probably right. the only thing i saw anandtech note is that with nforce3 boards Tras of 10 provided the most consistent over different speeds, not the best. Its best that you find that sweet spot though.
 
Qwertyman said:
well your probably right. the only thing i saw anandtech note is that with nforce3 boards Tras of 10 provided the most consistent over different speeds, not the best. Its best that you find that sweet spot though.
Click to expand...

Yep, and that sweet spot varies based on so many variables, it's almost unique for everyone unless they have 100% identical hardware and requirements.
 
My mem/mobo combo is Corsair 4400C25PT and DFI NF4 Sli-D. The optimum tRAS seeting on my system (for bandwidth) is 10 though I haven't determined yet how stable this is. Seems to vary depending on the mobo and memory you use.
 
I have pretty much the same sytem that anandtech used before the NF4 came out, and he still said to go for 10.. I guess it can vary down as far as specific modules. Just seems strange that all the different bench's out there give conflicting results..

wish there was something that was solid enough to be the defacto reference bench..
 
Hey, this is part of what makes overclocking an art. It's like modding cars. I might add a bit more boost to my turbo, while another guy might focus on advancing the timing a little. There are lots of variables in tweaking, and yes each new chipset will probably have it's own 'sweet spots'.
 
Unfortunately, tRas (timing Row active strobe) is not an exact science. While tRas=10 works for most people at tCl=2(Cas), it doesn't work for everybody. It would depend largely on your CPU's memory controller, as not all memory controllers are made equal... that and the tCl latency, of course.
 
You must log in or register to reply here.
Back
Top