Quick question about DDR4 2133 vs 3200

WAZ

Golden Member
Jan 17, 2001
1,642
2
81
I was about to grab 32GB of DDR4-2133 for my new Gigabyte Z170X

Then I found that for the RAM I was looking at (Corsair Vengeance), I could get DDR4-2666, which I read was the "sweet spot" for Z170 right now, or even DDR4-3200 for only like $20 bucks more.

Money-wise, seems like I should just the highest one I can, right? But if I do, do I need to get into overclocking to even use it at those speeds? Sorry, I've never OC'd before... I am building with an i5-6600K, so I CAN and should overclock, and I'll read up on that separately. But just needed to know if I can just plug-n-play with 3200 (the motherboard is spec'd for 2133 by default, but allows for the higher speeds as well) or if there will be more of a "process" to using the higher speed RAM.

Thanks!
 
Feb 25, 2011
16,790
1,473
126
The memory controller is part of the CPU - the motherboard manufacturers can't really guarantee anything the CPUs aren't technically capable of.

Browsing Intel's Ark site, it looks like the 7700K only supports up to -2400, and most other 7xxx series CPUs top out at DDR-2133. So I suppose from a "technically correct" standpoint, anything over that is "overclocking."

The spec list on the motherboard's support page seems to indicate anything over 2133 is considered overclocking by Gigabyte, also.

The Gigabyte website includes a supported memory list for that motherboard, with modules up to DDR4-3600 speeds. You would presumably have to enable the appropriate XMP profile as outlined in the manual.

In general, the difference in speed between the midrange (2133-2666) and the high end (3200-3600) is something that you can spot in a synthetic benchmark (although it's not usually linear because of higher latency at higher speeds), but which will have very little impact on day to day usage or the performance of most applications. So I wouldn't worry about it.

In general, I'd rather have slower RAM with a better warranty from the qualified hardware list, than faster RAM without. If the more "cautious" buy also happens to be cheaper, than you can spend that money on a bigger SSD or something instead.
 
Last edited:
  • Like
Reactions: WAZ

WAZ

Golden Member
Jan 17, 2001
1,642
2
81
Thanks! :) I did see that the -2133 ran at a lower latency (CL13) than the others... -2666 thru -3200 look like they're all CL15 or 16.

I suppose if the difference is negligible anyway, the lower latency is a plus. Though this may make going for the 6600K for overclocking kinda irrelevant.
 

UsandThem

Elite Member
May 4, 2000
16,068
7,380
146
I have a similar Gigabyte board, and the one thing to keep in mind if you want to use 32 GB of RAM. On my board's QVL, and using Corsair's and G Skill's "memory finder" tools, there is only one Corsair 8 GB DDR 3200 module certified to work properly on board. In fact, G Skill doesn't recommend any 8 GB modules above DDR4 3000 for my board.

There are even 8 GB DDR4 3000 modules from Corsair and G Skill that only work properly when limited to two modules. Filling all four slots caused some type of issue in the various certification tests.

I run DDR4 2133 because I bought it when prices for faster modules were much more expensive, but if I were buying today with the small price difference, I probably would pick up DDR4 3000, but I also don't run 32 GB. Depending on what memory review you believe, some show a noticeable difference, while some reviews say there is hardly any difference. It has been a hot-button topic around the various computer forums for a while. I am in the camp that it doesn't matter that much for most things (outside of using integrated graphics).

When all four memory slots are populated with faster RAM, it is tougher on the memory controller than just using RAM in two slots. But like Dave mentioned, it all comes down to the "silicone lottery" and what your particular CPU's controller can handle. Some are better than others.
 
  • Like
Reactions: WAZ
Feb 25, 2011
16,790
1,473
126
Good point. Yeah, 4 DIMMs on a motherboard will almost always be limited in speed vs. 2 DIMMs.

Thanks! :) I did see that the -2133 ran at a lower latency (CL13) than the others... -2666 thru -3200 look like they're all CL15 or 16.

I suppose if the difference is negligible anyway, the lower latency is a plus.

Latency as a CL number is relative to the clockspeed.

Basically, your RAM will function at a certain number of nanoseconds, period. What combination of clock speed and CL settings you use to get there is up to you.

CL16 at -3200 is about the same latency (in nanoseconds) as CL13 at -2133. (My math is probably wrong, but you get the idea.)

Though this may make going for the 6600K for overclocking kinda irrelevant.

I don't see why it would. It's very rare that an application is memory-bottlenecked (programmers are generally careful about that) so faster CPU clocks tend to correspond pretty well to faster overall system performance. I'm not as bullish on overclocking as I used to be, but if that's your bag, crank the clocks speed up and let your freak flag fly.

I would be looking at 7xxx series chips now, though. Just because why buy last week's stuff?
 
  • Like
Reactions: WAZ

BonzaiDuck

Lifer
Jun 30, 2004
15,727
1,456
126
I had an exchange with G.SKILL's tech-support, who are always good with their e-mail responses. I'll take a step backward to explain.

In the last DDR3 generation of RAM, you could, if you wished, fill your motherboard with 4x8GB of their "OC" XMP RAM with 2x 2-module kits and get the full XMP spec performance. I think this may have worked with lower-tier motherboards like the ASUS "-A" model-line. We have a Z77-A motherboard, and you could mix the size of the two kits, so we put in a 2x4 and a 2x2 kit for a total of 12 GB.

So I was fishing for a similar way to add RAM to my Skylake, which as a 2x8GB set of DDR4-3200 XMP TridentZ 14-14-14. I couldn't find anything like a 2x2GB kit or a 2x4 kit, and I put in a tech-support query asking for a recommendation. My past history with G.SKILL leaves me trusting that they're not just "trying to sell more stuff," and their recommendations have always been spot-on.

They said that adding a second 2x8GB DDR4 kit might make it difficult to run at the XMP spec, and they'd have to be set lower -- either by reducing the speed, or loosening the timings from 14 to 16. So they recommended simply replacing the 2x8 kit with the identical 2x16GB=32GB kit.

I can't speak to the Gigabyte Z170X, but you can cross-reference the board's QVL list and a RAM-manufacturer's "configurator" link. As for Corsair, I've got a set of 4x4GB DDR3-1600 which are stellar. I picked them up as part of someone's used bundle. I'd used prior-gen Corsair at least a couple times.

However, this is the way you do it. The board will, by default, choose 2133Mhz as the speed for a set of DDR4-[higher speed]. You should overclock your processor at that speed first, then choose the XMP RAM profile for the spec. There shouldn't be a problem with it proving stable right away, as far as my G.SKILL experience goes. Of course, you'd have to stress-test again. If there is any problem with the stressing, it's more likely that it needs more VCORE. Upping the VCCIO (IMC) wasn't necessary for me with my 2x8 kit.

Personally, I don't think it's worth the trouble "OC'ing" the RAM beyond its "OC-XMP-spec." Just buy the speed you want and run them at their XMP specification.

PS . . . If you purchased RAM at a spec higher than 2133, they should run at the spec (3,000 for instance) out of the box at stock CPU setting using the RAM XMP profile. And as I said, I only had to bump up my VCORE a tad to get complete stability with the RAM XMP profile while OC'ing the CPU.

There are a lot of different views concerning the CPU OC voltage. I see people who say they run their Skylake at 1.38V to get 4.7 or 4.8. I have a different view, based on an engineering white-paper I found about the processor. For an i7-6700K (not sure about 6600K) you should be able to find the perfectly-stable setting by starting with these numbers for each OC going up the scale:

4300 . . . 1.20V
4400 . . . 1.232V
4500 . . . 1.280V
4600 . . . 1.344V
4700 . . . 1.408V

You could either start at numbers close to but below these and test with possibility of bumping up the voltage, or you could start with the numbers and test by lowering the voltage.

if it's a 6700K, the processor will either run at 4600 or 4700 with stellar cooling -- or it won't. Luck of the draw. Another member got his 6600K to 4.5 Ghz, and I'm not sure about his voltage, but I think it was around 1.28. He was rocking a Hyper 212 EVO cooler, and his temperatures reached 80C.

If you reach that temperature with a stable overclock, you may not get much further unless you can improve the cooling. Heat causes electrical noise. And the Skylakes are more sensitive to temperature than earlier models. But they're supposed to be more resilient to voltage degradation -- how much, I cannot be sure. It could be around 1.4V, though.
 
Last edited:
  • Like
Reactions: WAZ

energee

Member
Jan 27, 2011
55
2
71
I was about to grab 32GB of DDR4-2133 for my new Gigabyte Z170X

Then I found that for the RAM I was looking at (Corsair Vengeance), I could get DDR4-2666, which I read was the "sweet spot" for Z170 right now, or even DDR4-3200 for only like $20 bucks more.

Money-wise, seems like I should just the highest one I can, right? But if I do, do I need to get into overclocking to even use it at those speeds? Sorry, I've never OC'd before... I am building with an i5-6600K, so I CAN and should overclock, and I'll read up on that separately. But just needed to know if I can just plug-n-play with 3200 (the motherboard is spec'd for 2133 by default, but allows for the higher speeds as well) or if there will be more of a "process" to using the higher speed RAM.

Most likely that DDR4-3200 kit achieves its rated speed at a voltage that exceeds the JEDEC standard of 1.2 V -- probably 1.35 V.

My advice if you want to minimize headaches: buy Crucial or Kingston brand RAM that is designed to run at the standard 1.2 volts.

Of all 288-pin DDR4 UDIMMs currently available from retailers, only two brands to my knowledge have anything to do with actually making memory (i.e., manufacture of chips and modules): Crucial (Micron) and Kingston. Crucial/Micron makes their own memory chips as well as modules, whereas Kingston only makes the modules. Both have an excellent reputation and Kingston is currently the largest DRAM module supplier in the world. I would add Samsung to the list but I don't see availability through retail channels.

Quality memory should run at its rated speed whether its purchased as single modules or a quad-channel kits. It's the responsibility of the memory manufacturer to provide modules with sufficient overhead for proper operation in any supported configuration. Some companies very aggressively bin memory modules and that's typically the source of the problem when users report difficulties adding additional modules to their system.

If you want a truly "plug 'n play" solution, you can buy 32 GB of Kingston Hyperx Fury DDR4-2666 RAM. It should run at its rated speed (@1.2 V) without the need to change any BIOS settings. Basically, Hyperx Fury saves a minute or two of your time by removing the need to enter BIOS setup and manually select the XMP profile. That's because the faster speed is already programmed into the SPD section of the EEPROM.
 
  • Like
Reactions: WAZ

WAZ

Golden Member
Jan 17, 2001
1,642
2
81
Thanks for the advice everyone. I think I'm going to cancel my $186 order for 32GB of RAM and just go with 16GB. Then I can put that extra $100 towards exchanging my i5-6600K for an i7-7700K.

Looks like for almost the exact same $450 price, give or take a few bucks, I could get an i7 ($350) with 16GB of DDR4 ($100)... which should trump an i5 ($250) with 32GB ($200).

I'll probably stick to -2133 or -2666 since those seem to be about the same price. I probably won't bother paying $20 more to push for -3000+ since the performance payoff doesn't appear to be worth it.
 
Last edited:
  • Like
Reactions: UsandThem

Dasa2

Senior member
Nov 22, 2014
245
29
91
in the game tests i have done the fps increase from 3200c14 to 3866c16 is greater than the jump from a 6700k at stock vs overclocked to 4.7ghz core\cache
imo its well worth the small price for 16g ~3000c14\c15 which should be fairly easy to run on any half decent mb

3200c16 has a much lower final latency than 2133c13
although sub timings can have a fair effect to
2666c16 would be close to 2133c13

16g vs 32g i5 vs i7 really depends on the software your using
32g makes no difference if you don't need it but it can make all the difference if you do
ht is the main advantage to a i7 and if the software doesn't utilize it then its no faster
 
Last edited:
  • Like
Reactions: WAZ

WAZ

Golden Member
Jan 17, 2001
1,642
2
81
I'm looking primarily for games (Fallout 4, Hitman, Battlefield 1, etc), and am also a graphic designer who works in Photoshop (mostly web, so rarely HUGE images) and some light video editing work. I'm sure 32GB of RAM could be helpful for that, but I don't think I'm in a situation where the extra 16GB would help me more often than the i5 to i7 jump will much most of the time.

That said, I did find DDR4-3000 for about $15 CHEAPER than -2133 (same exact brand and model -- Corsair Vengeance LPX for $97), so I decided to get 16GB of that along with the i7-7700K. I think that's a better use of the $450 budget for my needs. :) I also figure I can more easily add 16 more GB down the road if I need to, easier and cheaper than swapping out my processor.
 

Dasa2

Senior member
Nov 22, 2014
245
29
91
sounds like 16gb should be perfect for your needs and as you say you can always add more
the i7 will be of some help at times so i think you made the right choice there

as for fallout 4 its one of the games i tested
its gpu limited in some parts but in the cpu limited parts of the game this is how it responds to higher ram\cpu speeds
fallout4%20cpu%20vs%20ram.png

its worth mentioning that the minimum fps are more accurate than the average in this test
 
  • Like
Reactions: WAZ

BonzaiDuck

Lifer
Jun 30, 2004
15,727
1,456
126
Between the 3200 14-14-14 and the 4000 c17 models, 2.7 to 4.5 fps gained out of a base just short of 70 isn't a deal-maker for me.

Let's suppose that I now felt it reasonable or desirable to have 32 GB instead of my original 16. G.SKILL informs me they recommend simply replacing one 2-module kit with another, and that running 2x kits even of the same model DDR4-3200 cannot be guaranteed to run at the full XMP spec. This is a new twist as opposed to the DDR3 regime, for which 2 kits of even different size but same model will run at the same spec but limited to CR or T = 2.

Suppose the new 2-module kit for the full 32 GB costs me $250. The single kit of identical 2x8 16GB is likely ~ $130.

How much would I give up in performance to use four slots at 3000 14-14-14 as opposed to two slots @ 3200 14-14-14? Consider the savings up front is about $120 -- using today's prices. From these benchtest results shown by Dasa2, it would seem "not too much."

Another question might be "How much additional stress on the IMC for 4-slots of RAM, and would that matter? I can tell you it doesn't for Sandy Bridge and DDR3. I could wonder about Skylake and DDR4, because I know damn well I'll have to twist up the VCORE a tad just like I did with my Sandy.
 

WAZ

Golden Member
Jan 17, 2001
1,642
2
81
Yeah, I've seen benchmarks which show little to no improvement whatsoever going from -2133 to -3200. Then others like Dasa's which show at least a small gain in fps in some games. A 5-10 fps increase isn't worth a ton of money, but I figured for like a $10 difference, if you can find a good deal on -3000 or higher, then might as well go for the faster. Worst case, it scales itself back to -2133; best case the XMP profile gives me a nice little boost, and in the future if I OC my 7700K, I could see a small fps improvement in games like Fallout. Again, not enough to be worth a much bigger investment... but for the cost of one less trip to Taco Bell, I figured the faster RAM would be worth it. :)

Amazon is going to be pissed at me for all the parts I've canceled and/or returned, but I did finally settle on 16GB of G.SKILL TridentZ 3200... $109... Good deal for $10 more than the set I originally had. And that left me with the extra money from returning my 32GB that I could get the i7-7700K. Just had to finally commit and start building.
 

Dasa2

Senior member
Nov 22, 2014
245
29
91
any games that show no difference from faster ram will also show no difference from a faster cpu due to a gpu bottleneck
some games will see bigger gains from ram than others with open world games seeming to have the most to gain

although it may only be a 5-10fps increase in the 30-40fps range thats a large % increase if it was at 100fps that would be ~20-30fps
similar to the difference between sandybridge and karby lake at the same clock speed

that kit sounds decent value hope your happy with it :)
 
Last edited:

BonzaiDuck

Lifer
Jun 30, 2004
15,727
1,456
126
Over some range of speeds, I think the price differential between a DDR4-2133 kit and maybe as much as 3200 is still chump-change when you consider that this is an investment that should last -- say -- 5 years.

My problem is when and how or even whether to decide for replacing the 2x8 with a 2x16, or getting another 2x8 and clock both kits down to 3,000 at the same timings. I'm only guessing that the VCORE of all slots filled will need a bump greater than it would by inserting a 2x16 kit. That's a problem of direction in the upgrade tree.

The second problem involves the "when" of replacing my experimental 250GB 960 EVO caching drive with a 1TB Pro or 1TB EVO. Which should it be: 1.2 Petabyte life expectation, or 400 TB guaranteed life expectation? Because the 300 MB/s difference in the sequential read spec is only a 10% difference, but the price-saving is maybe $180 at today's Egg prices -- roughly -- or 27% saving over the Pro.

Putting in context, it's all part of a wider plan and an annual budget for the 2017 year. And the experiment with the 250GB EVO is indeed a success. As far as needing more RAM as I continue to install software and test this system, I may have 6GB total of "standby" and "free" RAM -- "in-game," with other programs open and running.

Again, this is using PrimoCache and 4,096GB of my RAM kit for the caching, together with a ~40GB NVMe cache for my SATA SSD boot-disk and ~65GB for my SATA HDD, spinning at the lollygag rate of 5,400RPM. After three days running, and also after running an initial local backup against an uncached hot-swap drive for the SATA SSD and HDD, both caches have filled up leaving only about 8 GB in each one.

It just gets back to the issue of need, if this approach is as successful as I think it has been. And I continue to be polite and civil with naysayers, because they have important things to point out.
 

Dasa2

Senior member
Nov 22, 2014
245
29
91
the mb can have a large impact on how much extra v is needed for 2 sticks vs 4 but mostly its the v to the memory controller that needs to be upped system agent and vccio
the asus hero\gene for example have a thicker pcb than many cheaper mb allowing for better spacing and less crosstalk when running 4 sticks of ram
two 16g sticks can also be a bit harder on the memory controller if its double sided\dual ranked dimms vs single but i have no idea which is harder to run
 

BonzaiDuck

Lifer
Jun 30, 2004
15,727
1,456
126
the mb can have a large impact on how much extra v is needed for 2 sticks vs 4 but mostly its the v to the memory controller that needs to be upped system agent and vccio
the asus hero\gene for example have a thicker pcb than many cheaper mb allowing for better spacing and less crosstalk when running 4 sticks of ram
two 16g sticks can also be a bit harder on the memory controller if its double sided\dual ranked dimms vs single but i have no idea which is harder to run

I won't dispute the need to bump up VCCIO or VTT whatever the mobo calls it -- the IMC voltage. It simply cannot exceed 1.2V, which should hold true for DDR4. But with an overclocked system, I found that the sweet-spot for voltage had changed between 16 and 30mV -- for the VCORE. That was for using all four sockets of a DDR3 system, and if the need isn't as apparent with a Z170 chipset, I'd still want to test.

You'll certainly find out, whether you give it a shake-down stress-test or simply attempt to use it over a few days. Better the shake-down cruise.

I have the Sabertooth Z170 S, shouldn't be different than the Mark 1 but for lack of the duct plate and incidentals. It's a very solid board. The one with the 5-year warranty and the accommodation to "MIL_SPEC."

The risk of buying a second kit of the same model and spec should be limited to running all four modules at the full 3200 XMP specification. If I can't run them at 3200, I could probably set them to 3000 and the same timings. The cost would be maybe $140.

Instead of risk, I have the extra expense of around $270 for the 2x16 GB kit, and a pair of modules I'd either sell or put in the parts-locker.

Of course, the longer I put off my alternative plans, the less the extra change will matter. But bundled together with my OCD techno-lust for a 1TB 960 Pro or EVO, and a temptation to get a 27" 1440 BenQ "ZOWIE" monitor, I'll want to make the purchases in increments of $500+ each so I can take my time with no interest, while still not keen on letting go of the purse-strings quite yet. . . .

On the other hand . . . . and I can post in the appropriate forum for that, anyone interested in TridentZ flawless 16GB 3200's could PM me.