[XbitLabs] Analysis of Memory Frequency/Timing on IVB

[Termie]

Link: http://www.xbitlabs.com/articles/memory/display/ivy-bridge-ddr3.html

Interesting test, and one I don't see a lot of reviewers focusing in on. IVB really takes advantage of higher frequencies, and it seems its worth it to push for those frequencies even if you have to relax timings. Seems like the sweet spot is between 2133 and 2400.

The ironic reveal at the end is that the $200 set of G.Skill memory that they test with is probably identical to the $50 Samsung set available at Newegg.

Samsung ($50): http://www.newegg.com/Product/Produc...82E16820147096
G.Skill ($200): http://www.newegg.com/Product/Produc...82E16820231591

 

[Prey2big]

How come the Ivy Bridge-platform show substantial gains from faster memory but the Sandy Bridge-platform only show very marginal gains?

 

[ElFenix]

interesting results compared to sandy. though the article states no real changes to the memory controller there is an obvious latency improvement for IB.

they did have to run the memory out of IB spec to get it that fast.

 

[ShintaiDK]

The numbers looks pretty consistent with SB cept gaming. But I´m not sure I put that much faith into xbitlabs since those numbers are all over the place.

So as I see it, its still the exact same as SB.

 

[taltamir]

How come the Ivy Bridge-platform show substantial gains from faster memory but the Sandy Bridge-platform only show very marginal gains?

This is explained in the conclusion of the article:
http://www.xbitlabs.com/articles/memory/display/ivy-bridge-ddr3_7.html#sect0

Basically, the improvement per step remains the same, but what changed is that intel improved the memory controller to allow it to use RAM that is much faster, so much faster that on sandy bridge you couldn't use it even if you overclocked (it will not be stable).

 

[utahraptor]

Sounds like the 2400 Trident 8 gig modules I have my eyes on will help

 

[cmdrdredd]

This gave me an idea.

Samsung Low Profile 30nm @ 1.56v

 

[Termie]

This gave me an idea.

Samsung Low Profile 30nm @ 1.56v

Exactly...since that's basically the same kit they're using.

Can you post the same benchmark at 2133, which is what you typically run? Xbit argues that the tighter timings you would have wouldn't benefit you much, if at all.

 

[cmdrdredd]

This is a result I have from 4.6Ghz

 

[Aikouka]

Didn't Anandtech run a similar test? I can't remember if it was on Sandy Bridge or Ivory Bridge, but I'm pretty sure that they also found that bus speed > timings in regard to overall performance increase. That article sure made me think twice about worrying so much about some of the low CL RAM.

 

[Magic Carpet]

Not Found

The requested URL /articles/memory/display/ivy-bridge-ddr3_7.html was not found on this server.

Somebody must have pulled the plug

 

[Makaveli]

This is a result I have from 4.6Ghz

did you drop the memory timings to get to 4.6?

All your numbers are lower and its shows memory latency is up in this screenshot?

I also see bus speed is lower which may be the reason.

 

[boxleitnerb]

Article not found?

 

[Termie]

did you drop the memory timings to get to 4.6?

All your numbers are lower and its shows memory latency is up in this screenshot?

I also see bus speed is lower which may be the reason.

What he's showing is the benchmarks at 4.5/2400 and 4.6/2133. Not apples-to-apples, so it's hard to make a direct comparison, but the 2400 memory speed, despite slightly looser timings, has better overall results in the memory benchmarks.

Article not found?

The XbitLabs website is down. I'm sure it will be back up eventually.

 

[Makaveli]

What he's showing is the benchmarks at 4.5/2400 and 4.6/2133. Not apples-to-apples, so it's hard to make a direct comparison, but the 2400 memory speed, despite slightly looser timings, has better overall results in the memory benchmarks.



The XbitLabs website is down. I'm sure it will be back up eventually.

Makes sense now thanks for the breakdown

 

[cmdrdredd]

did you drop the memory timings to get to 4.6?

All your numbers are lower and its shows memory latency is up in this screenshot?

I also see bus speed is lower which may be the reason.

The first shot I posted is 4.5ghz ddr3-2400 10-11-11-28 @1.56v. The second is from an old test at 4.6ghz ddr3-2133 10-10-10-28 @ 1.47v

I can run either CPU overclock but 4.5 is my 24/7. The numbers for memory are lower because 2400 > 2133

 

[cmdrdredd]

I can run 4.5 benches at 2133 when I get home if you want them.

 

[Termie]

I can run 4.5 benches at 2133 when I get home if you want them.

It would be interesting to see.

Now that you've done these tests, are you going to stick with 2400 or go back to 2133?

 

[Rvenger]

Really curious to see as well.

If this is the case, I will sell my 16gb 1600mhz RAM and buy 8gb of 2400mhz.

 

[cmdrdredd]

It would be interesting to see.

Now that you've done these tests, are you going to stick with 2400 or go back to 2133?

2400 wasn't memtest stable at 1.56v, but I could bench and play a few games that way (not extended periods). Gave errors. Need 1.6v or more I think. I managed to boot at DDR3-2600 13-13-13-35 CR2 at 1.63v but I do not like that voltage at all.

edit: tried to screenshot 2600 but BSOD lol...not stable even 1.65v doesn't help. I could probably get 2400 stable but I am not sure how 1.6v+ is for long term. Like to stay lower if possible.

I'm back to DDR3-2133 10-10-10-28 CR1 1.475v

 

[Termie]

2400 wasn't memtest stable at 1.56v, but I could bench and play a few games that way (not extended periods). Gave errors. Need 1.6v or more I think. I managed to boot at DDR3-2600 13-13-13-35 CR2 at 1.63v but I do not like that voltage at all.

I'm back to DDR3-2133 10-10-10-28 CR1 1.475v

Assuming for argument's sake that your Samsung modules are in fact the same as the G.Skill Trident set that Xbit used, you might try the timings they benched at:

•DDR3-2133 11-11-11-33-1N;
•DDR3-2400 11-11-11-33-1N;
•DDR3-2667 11-13-13-35-1N.

The Tridents are rated at:
•Nominal frequency: 2600 MHz;
•Timings: 10-12-12-31-2N;
•Voltage: 1.65 V.

It will definitely be hard to do that with under 1.5v, though, so I understand your concern, but you might not need much more than that if you keep the timings loose, which was the aim of the article. It looks like you were still trying to keep them as close to the limit as possible.

 

[cmdrdredd]

DDR3-2400 1.56v 11-11-11-33 CR1

DDR3-2666 1.59v 11-13-13-35 CR1