Is upgrading memory worth it? (MHz)

[It's Not Lupus]

I have a 2500K (4.4Ghz) and this memory:
http://www.newegg.com/Product/Produc...45&ignorebbr=1

and this motherboard:
http://www.asrock.com/mb/Intel/P67 Extreme6/index.asp?cat=
which supports up to "Dual Channel DDR3 2133(OC)"


Would I notice any increase in FPS in games with 1866MHz or 2133MHz memory instead of 1600MHz? I ask because I might transfer the old memory to another system which could use an increase in capacity.

 

[BonzaiDuck]

I have a 2500K (4.4Ghz) and this memory:
http://www.newegg.com/Product/Produc...45&ignorebbr=1

and this motherboard:
http://www.asrock.com/mb/Intel/P67 Extreme6/index.asp?cat=
which supports up to "Dual Channel DDR3 2133(OC)"


Would I notice any increase in FPS in games with 1866MHz or 2133MHz memory instead of 1600MHz? I ask because I might transfer the old memory to another system which could use an increase in capacity.

A lot of our "knowledge" here at the forums is the spawn of consensus and information passed on by other users.

The consensus seems to be this. For either the Z68 or P67 motherboard and particularly the Sandy Bridge K processors, anything beyond 1866 is of little value. Even the performance gain from 1600 to 1866 is somewhat insignificant.

First -- What type of RAM are you using? My original G.SKILL RipJaws DDR3-1600's (9-9-9-24-2N and 1.5V) could be easily overclocked to 1866 @ 10-10-10-32. I was able to run two kits of the same (all four slots filled) with that RAM. I replaced them unnecessarily, knowing that two is better than four and trying to resolve an occasional instability. It turned out that the instability had other causes: over-allocation of Z68 motherboard resources, and possibly an interaction of certain software and hardware.

Also, with either the RipJaws or RipJaws-Z modules, you can get a better performance boost by running them at a command rate of 1 instead of 2. You would only need to bump up the VCCIO a couple notches, or raise the PLL Voltage if you'd set it initially as low as 1.65V.

You may have other RAM like Corsair, Crucial or Samsung. In my case, G.SKILL tech support confirmed the viability of the command-rate setting. I remember some years back that Crucial lambasted me for using CMD=1, in an exchange with them over an RMA request. But those were DDR2. I wouldn't know for sure what options you have.

Even so -- you might want to investigate. For instance, make a point of adjusting the BIOS for "XMP," while assuring the speed is set to the RAM spec (in your case, 1600). Set all the RAM timings to "Auto" including the command rate. Then reboot and see if the BIOS shows the modules running at CMD=1.

UPDATE: Ho-ho! Ha-Haaah! I overlooked your first link! Yes! Those should do nicely! What I'd try? See if you can get them to run stably at 1866 and 9-9-9-24.

I'd had another exchange with their Tech-support, who pretty much confirmed that different model kits likely would clock the same by adopting the higher-rated RAM's timings. THEN! Try the command-rate thing. I'm guessing you can still run those at 1.5V and 1866!!

But if you can't find the proper (looser) timings for 1866, I think you're still guaranteed command-rate of 1 with some minor tweaks -- if necessary.

Just be sure you test the new settings thoroughly before assuming your system will be totally stable afterward. I'd run HCI Memtest-64 for at least 300% coverage, or three 100% iterations.

ANOTHER UPDATE: I think you can set the BIOS to "XMP" instead of "manual," then choose the higher speed and attempt to leave the latency settings on "Auto." See what latencies that gives you before proceeding. Especially -- XMP (again confirmed by G.SKILL tech-support) will auto-adjust the "advanced" timings with more aggressive settings. Otherwise, see if you can't get 1866 with "XMP," the desired speed, and manual timings. Leave the advanced timings set to "auto."

AND YET ANOTHER [UPDATE]: I'm wondering why you can't run that 2500K at 4.5 instead of 4.4. What settings are you using? We can continue this ancillary discussion at your choice and leisure.

 

[BrightCandle]

Tests show only a few particular programs benefit from more RAM bandwidth, its a complicated subject really. Most games don't show much benefits but tests done in the Arma 3 community show that Arma actually scales incredibly well with RAM bandwidth and its worth getting 2133 or above for that one game. But generally most games don't seem to benefit. Then there are the compression programs which gain a bit with faster RAM and in the past some of the video encoders did as well but not sure if that is still the case.

So in limited circumstances there can be a small improvement in performance. If the cost isn't too much higher then its worth getting a higher speed RAM for those particular circumstances.

 

[BonzaiDuck]

Tests show only a few particular programs benefit from more RAM bandwidth, its a complicated subject really. Most games don't show much benefits but tests done in the Arma 3 community show that Arma actually scales incredibly well with RAM bandwidth and its worth getting 2133 or above for that one game. But generally most games don't seem to benefit. Then there are the compression programs which gain a bit with faster RAM and in the past some of the video encoders did as well but not sure if that is still the case.

So in limited circumstances there can be a small improvement in performance. If the cost isn't too much higher then its worth getting a higher speed RAM for those particular circumstances.

Yeah -- but I think it also depends on the processor. For the SB-K's, I thought 1866 was as good as it gets. IIRC, it depends on the integrated memory controller, and I think the Ivy processors were better for that aspect.

Unless he's really eager to relocate those G.SKILLs to another system, I don't think he really "needs" to spend the money. [Just like I didn't "need" to buy the Ripjaws "Z" kit.] IF he wants to? That's something else. He could downclock a set of 2133's if he wanted. Or he can overclock what he has. I'm just pretty sure that 1866 was about as good as it gets for the SB IMC. Maybe you're right though. How would it depend on any difference between an SB-K and an SB-E, such as yours?

 

[It's Not Lupus]

...

I never thought about overclocking my memory for this system until now

The XMP profile sets the memory to 1600 only so I manually set it to 1866. On auto for those four timing settings after a reboot, the values were set to, I think, 13-13-13-33 which I thought was high compared to settings of retail kits.

I misread your post and tried 1866 and 8-8-8-24 which the system did not like and wouldn't even POST. At 1866 and 9-9-9-24-1N, system passed one round of memtest86. I'll do more passes later and look into HCI Memtest-64.

I don't remember why I left it at 4.4GHz but 4.5GHz was also stable so I changed it.

I might still buy more memory on Black Friday for the other system.

 

[dave_the_nerd]

If you can OC it, sure, why not? Upgrading RAM speed if you're already at DDR3-1600 isn't worth spending money on though.

 

[BrightCandle]

Yeah -- but I think it also depends on the processor. For the SB-K's, I thought 1866 was as good as it gets. IIRC, it depends on the integrated memory controller, and I think the Ivy processors were better for that aspect.

The test I saw was run on SB and SB-K, and it showed decent scaling all the way up to 2400mhz which is where the users memory stopped going faster. But its not the rule that this is what happens, most games don't benefit from large memory bandwidth and neither do most programs. You are quite right that 1600-1866Mhz is the sweet spot for many of the benchmarks that sites run.

 

[_Rick_]

for many of the benchmarks that sites run.

Of course, many of the benchmarks that are run, don't stress the memory subsystem very much.
I wish developpers would profile their games better and give a sot of "GPU/CPU/RAM/bandwidth" ratio for some common configurations, so it becomes easier to determine which component is currently the bottleneck.

Especially if games ever get to the point, where they use multi-threading, you could have completely decoupled/asynchronous rendering processes which load up the GPU to the limit, but you could still be effectively limited elsewhere, and the AI/physics/whatever is actually simply not doing things during some frames.

Maybe someone should develop a calculator, where developers can enter some configurations and results, and given a certain hardware configuration, the app points to the most likely effective bottleneck, and recommends an upgrade. Of course, getting industry support for something like this is as close to impossible as it gets

 

[Cerb]

Especially if games ever get to the point, where they use multi-threading, you could have completely decoupled/asynchronous rendering processes which load up the GPU to the limit, but you could still be effectively limited elsewhere, and the AI/physics/whatever is actually simply not doing things during some frames.

We're already there (not the actual rendering stage, but setup for it is supported by nVidia, and quite a few games are pretty well multithreaded, as well). It's spreading to more and more games over time.

Some games do benefit from faster RAM by more than negligible amounts, it's just that the differences still pale in comparison to a higher CPU bin, mildly faster GPU OC, etc., making spending on faster RAM typically a last step in a heavily OCed high-end system, rather than a normal thing to do.

http://www.anandtech.com/show/7583/adata-xpg-v2-review-2x8-gb-at-ddr32800-121414-165-v/4

Compare the minimum FPS. Notice that 1600 and 1866 with lowish timings are right up there with 2400 and 2800MHz RAM, and the greatest difference measured was all of 5%, in one game. Check out other reviews at other sites, with more games tested, and you'll find that's about as high as it goes. In many games, it will be <1% (which is around the difference between motherboards, CPU and GPU samples, testing error margins, etc.), and rarely break past 3%, even with nearly double the MHz.

 

[BonzaiDuck]

I never thought about overclocking my memory for this system until now

The XMP profile sets the memory to 1600 only so I manually set it to 1866. On auto for those four timing settings after a reboot, the values were set to, I think, 13-13-13-33 which I thought was high compared to settings of retail kits.

I misread your post and tried 1866 and 8-8-8-24 which the system did not like and wouldn't even POST. At 1866 and 9-9-9-24-1N, system passed one round of memtest86. I'll do more passes later and look into HCI Memtest-64.

I don't remember why I left it at 4.4GHz but 4.5GHz was also stable so I changed it.

I might still buy more memory on Black Friday for the other system.

It failed to occur to me that memory sold as "DDR3-1600" might not have a "DDR3-1866" profile -- sorry. On the other hand (only looking at my own BIOS), the choice of available "Profiles" for XMP was one menu item, and there was a separate item for choosing DDR speed. You may look into that as inclined.

Try the looser timings and 1866 first under the command-rate "2N." Or -- try 1N under the DDR3-1600 profile.

Also, you mentioned excessively-loose timings. With my G.SKILL "GBRL" Ripjaws kits (DDR3-1600), I had to manually enter looser timings for 1866, originally sorted out and tested by another forum member here. The stock settings were 9-9-9-24 and the looser timings for higher speed were 10-10-10-32. What we mutually saw: with the looser timings, you did not have to set the vDIMM higher than 1.50V. So when I see timings of 8-8-8 for 1600 on a kit, well -- you know . . . . test those suckers as you've begun to do!! My "gerbils" passed HCI-Memtest-64 1000% coverage (four days!) at the higher speed and looser timings . . . Of course, I wasn't experimenting with "XMP" profiles when I was doing that: I'd entered all settings manually. You can still set the XMP profile and then manually tweak the speed and timings -- or so you should be able to do so.

UPDATE (AGAIN): See if this helps. On my motherboard, you can set "XMP" versus "Auto" or "Manual" for that particular configuration item. Now -- my RAM is spec'd and sold as Ripjaws Z DDR3-1866. But my motherboard has another item in the menu that includes "XMP": A drop-down of speed options for running the RAM. These would not be specified in the SPD or XMP profiles of the RAM; they are an option of the motherboard. The options in the drop-down menu go up to DDR3-2400 (maybe higher, but I remember that number). Such a number wouldn't be in the RAM's XMP profile, or it is more likely an option provided by the motherboard in conjunct with choosing an XMP profile.

If you're going to choose running the DDR3-1600's at 1866, make sure you also manually tweak the latencies. With the looser timings, there's a good chance you won't need to bump up the vDIMM or RAM voltage above 1.5, but if you do, the adjustment shouldn't need to be very much.