Trick with tRCD 2-2-2-6 versus 2-3-2-6

[stevejst]

Just noticed something rather surprising. tRCD relaxing from 2 to 3 makes a huge difference.

My settings on 2-2-2-6 Kingston HyperX, street racer, 5:4 with 270:216 (2.4C @3.243GHz) gives 2755 MB/s in Memtest86. Using 2x256 Kingston HyperX PC 3200, dual channel, 511 MB with 1 MB memory hole.
All the same except for 2-3-2-6 gives 2585 MB/s.

This is 6.6% difference in the bandwidth!
Similar difference is visible in Sandra.

No wonder Corsair and Kingston are now selling "faster" 2-3-2-6 memory.

 

[aggressor]

Why is this surprising? I'd expect 2\2\2\6 to be faster than 2\3\2\6

 

[stevejst]

Of course but I did not expect the difference to be 6.6%. That is huge.
For example relaxing for 1 on tRD that many overclockers are so fond of will give you only about 2 MHz on FSB. This is much bigger.

Or more to think about, BH-5 chips are capable running 2-2-2-5 on 400 MHz, they are not very likely to run 2-2-2-6 on 433 MHz but they are (most of them) able to run 2-3-2-6 on 433 MHz.
Think about that when you are buying Kingston HyperX PC 3500.

 

[edmundoab]

i know I had a slight performance drop when I changed by mem timings of 2-2-2-6 to 2-3-2-6
just wanted to see how much a difference it caused...

i didn't know it was 6.6% though...

 

[stevejst]

I just switched my GEIL Ultra Platinum PC 3500 to Kingston HyperX PC 3200. GEIL can run up to 484 MHz on 2.8V but cannot run 2-2-2-6 on 400 MHz. Kingston on the other hand can do 433 MHz on 2-2-2-5 but cannot go anywhere above 445 MHz, no matter what timings. So for 2.4C I found that Kingston gives me better bandwidth. If I have 2.8C, 3.0C, or 3.2C processor, I believe GEIL would be better.

 

[edmundoab]

how about Corsair?

 

[BDSM]

You know 2-2-2-11 is the fastest setting, right?

 

[boshuter]

Originally posted by: BDSM
You know 2-2-2-11 is the fastest setting, right?

i think a lot of people haven't discovered that yet..

 

[stevejst]

That seems to be working on some AMD boards but it has to be some compatibility trick because it does not make much sense otherwise. It does work on my NF7-S though.

 

[edmundoab]

yeah my settings are at 2-2-2-11 too
someone posted a thread telling us that, I tried it and I benchmarked a marginal performance speed improvement

 

[oldfart]

Originally posted by: BDSM
You know 2-2-2-11 is the fastest setting, right?

That does not apply to Intel rigs.

 

[BDSM]

Hey oldfart.

really?... I haven't seen any becnhies on intel on this at all

 

[oldfart]

You can't even set it to 11. 8 is as low as it goes. I've tested 5,6,7,8. 5 or 6 is fastest, 8 is slowest.

 

[stevejst]

The gain from 2-2-2-6 to 2-2-2-11 I get on NF7-S is not much and you ought to be concerned what it is in "real world" applications.

 

[BDSM]

oldfart. I c.. thanx

steve.. Well.. the difference is tiny.. but there is no reason to not use this setting if available.

 

[TerryMathews]

2-2-2-11 is only for nForce chipsets. Has to do with how the memory controller works.

I can vouch for 2-2-2-6 being fastest on 760MPX. Checked every setting from 6-11 and used AIDA32 benchmark each time.

 

[BDSM]

Okay.. I believe ya all.. It's just that someone explained this really neatly for me and I thought it made really good sense then.

Sorta like you are trying to read a book.. and the last value 5-11 would be how long you got to read each page.. and if someone slammed the book shut after five you'd have to look it up again. So 11 would be better cuz then you could finish your job before the page was closed.. If you see what I mean.. I dunno where that original text was.

 

[AkumaX]

can anyone do a benchie w/ 2-2-2-6 vs 2-2-3-6

 

[pspada]

Originally posted by: boshuter

Originally posted by: BDSM
You know 2-2-2-11 is the fastest setting, right?

i think a lot of people haven't discovered that yet..

Some of us have discovered (with nForce2 chipsets!) that 11 is actually slower than 8.

 

[Megatomic]

Originally posted by: pspada
Some of us have discovered (with nForce2 chipsets!) that 11 is actually slower than 8.

I wish there was some continuity with this. I'm sticking with 11 myself but everyone should certainly use whatever works best for them.

 

[Slammy1]

LOL. I did run 2-3-2-6 with my HyperX, but my FSB gain by relaxing to 2-3-3-7 more than offset the loss. Now going to 2.5-3-3-7, that's one I can't seem to justify in performance boost, but I can't get 5MHz more at that level. I need to go back to a 1:1, figure out just what this thing can do, I'm getting seriously CPU bound on the 5:4.

 

[helloguy]

Originally posted by: TerryMathews
2-2-2-11 is only for nForce chipsets. Has to do with how the memory controller works.

I can vouch for 2-2-2-6 being fastest on 760MPX. Checked every setting from 6-11 and used AIDA32 benchmark each time.

u mean 2-2-2-11 gives more bandwidth than 2-2-2-6 ???? why is this? I thought the lower the faster ????

 

[Megatomic]

Originally posted by: helloguy

Originally posted by: TerryMathews
2-2-2-11 is only for nForce chipsets. Has to do with how the memory controller works.

I can vouch for 2-2-2-6 being fastest on 760MPX. Checked every setting from 6-11 and used AIDA32 benchmark each time.

u mean 2-2-2-11 gives more bandwidth than 2-2-2-6 ???? why is this? I thought the lower the faster ????

The reason for this (source: Mushkin) is that when the tRAS is set too low errors in data transactions occur forcing the system to have to make the transactions all over again. The net effect of this is that less actual data is transferred per given amount of time. Therefore, with a higher tRAS, fewer transactions are required to transfer X amount of data.

Lower tRAS -> rapid data transactions -> many data errors -> many transactions required to transfer data.
Higher tRAS -> slightly slower data transactions -> far fewer data errors -> far fewer transactions required to transfer data.

 

[helloguy]

cool....

ok,i did some benchmarks juz now using sisoft sandra 2003.
here;s wat i got :-

166MHz(2-2-2-6)
2531MB/s,2380.3MB/s (average of 3 tests ran)

166MHz(2-2-2-11)
2498.3MB/s,2342.6MB/s(average of 3 tests ran)

this means that I've no errors in data transaction when using 2-2-2-6 ??

 

[Soulkeeper]

not neccesarilly

just that the performance increase is great enough that even if memory accesses have to be repeated on occasion they are still not enough to make your system perform lower than at the higher tras