Does Memory Matter?

[sm625]

If you're using a recent i3, especially Skylake, then get faster RAM because its not very much to step up from 2133 to even 3000. The second part is if you plan on playing a lot of Fallout 4, get faster RAM no matter your platform.

But how much sense does it make to combine an i3 with premium RAM, when you can just get an i5 and use the cheapest RAM?

 

[JoeRambo]

actual answer: no
http://techbuyersguru.com/gaming-ddr4-memory-2133-vs-26663200mhz-8gb-vs-16gb?page=1

5820K has 10MB of L3 and Quad channel. So it is getting the least out of memory bandwidth increases and less benefits from effective latency improvements versus 4,6,8MB L3 CPU's.

The way "low clock" Broadwell benefited from L4 cache is clearly indicative that there are benefits to be gained from faster memory. Being able to compete with big boys in games clocked 0.5Ghz less is no small feet.

 

[ehume]

Nice reviews. They tell us that all that stuff doesn't matter. Call it DDR4 2166MHz 8 or 16GB and call it a day.

What does matter is form factor. Using an AIO? Then your RAM can sport a fancy coxcomb and whatever the marketers dream up for their RAM.

Air cooling? Especially if you will be going with a 140mm push fan, you will want Low Profile RAM, maybe even Very/Ultra-Low Profile RAM.

With voltages as low as they are now, no RAM needs a coxcomb. That stuff is all marketing.

 

[MongGrel]

Still use GSkill Ripjaw DDR3-1333 here with the OC'd X5680, but it is running 7-7-7-21 1T

It seems to scale well with this old setup.

Interesting read, I like keeping informed on the newer stuff a bit at any rate.

 

[Dasa2]

actual answer: no
http://techbuyersguru.com/gaming-ddr4-memory-2133-vs-26663200mhz-8gb-vs-16gb?page=1

shame its a useless comparison of games performance from ram without also showing the difference with the cpu at stock clocks to see how much the gpu bottlneck is effecting the results
i wouldnt be surprised if with the cpu at stock clocks some those tests were a little faster than overclocked simply due to margin of error like this test

http://www.kitguru.net/components/cpu/luke-hill/intel-core-i7-6700k-i5-6600k-skylake-cpu-review/8/

as for what ram to get
3000 ÷ 15 =200 $80
3200 ÷ 14 =228.5 $120
3600 ÷ 16 =225 $160
so 3200c14 may be slightly quicker than 3600 16 going by that simple formula but the difference is so small it would be hard to measure let alone notice
http://www.anandtech.com/show/8959/...-3200-with-gskill-corsair-adata-and-crucial/2
3000c15 to 3200c14 could be worth maybe 100mhz higher oc on the cpu in game at a rough guess maybe not noticeable on its own but each small tweak can add up

 

[Majcric]

To echo what others have already said I would get the fastest affordable ram for a new build.

It appears the faster ram helps mostly when the cpu is a bottleneck.

 

[Azuma Hazuki]

So, it looks like the sweet spot is DDR4-3200/14 if OC'ing and DDR4-2666 if not...?

 

[Dasa2]

So, it looks like the sweet spot is DDR4-3200/14 if OC'ing and DDR4-2666 if not...?

2666 with timings as tight as possible may be best for x99 i dont know

2133 is all you can do without overclocking the imc a little

there may be some applications that benefit more from the extra bandwidth added by higher frequency ram than the lower latency provided by 3200c14 but 3200c14 seems like the sweat spot for z170 to me

 

[know of fence]

Yeah, for whatever reason, Skylake can take advantage of higher bandwidth RAM, in a few of the newest games. Fallout 4 is the one where it seems to make the most difference, IIRC.

I've seen this a couple of times, that essentially "Skylake likes fast RAM". But it's definitely a wrong thing to say! RAM latency and bandwidth matter as much or as little as they ever did, but ...

• we have extremely powerful GPUs and the dx11 standard that swamp and hammer CPU and memory more than ever, especially in artificially constructed tests (pairing a Titan X with an i3)

• increased number of multi-threaded games played on a dual core CPU significantly benefits from RAM, which begs for a juicy and satisfying technical explanation, which I would like to hear.

Anyway, Skylake "likes" fast RAM as much as Haswell does.
I even have proof: just copy/paste the Eurogamer table into a spreadsheet. If you compare how the results change from one RAM speed to the next, there is barely any difference between improvements. In the last column I multiplied the numbers by (3.7/3.4) or the clock difference, assuming bandwidth related improvements would also increase with a higher cpu frequency.

Possible flaws: The clock gaps between the modules used is different, and we can assume that all 4 RAM types have different latencies/performance index. Still it's the closest we got to a fair comparison, yet.

 

[frozentundra123456]

Well, to me the answer is that "ram matters" for both. And even though this seems like a lot of games, Ryse and Crysis 3 seem like outliers, so I dont know if there is really enough data yet to draw a conclusion. In any case, considering the cost of a system, spending the extra what, 50 to 100 dollars to get reasonably fast ram and a >10% improvement for either cpu seems like a no-brainer. Now whether Skylake benefits more than Haswell, that is a harder question.

 

[know of fence]

@frozentundra123456
+10% improvement under artificial conditions, mind you, a permanent bottleneck caused by a 1200$ GPU "deep learning" GPU running 1080p resolution on a crap core-i3 (Eurogamer)
+2-3% under sensible conditions and a dual channel high-end setup (TBG/Termie's test) and
_less still for quad channel CPUs that mostly benefit from improved latency (TBG/Termie's test)

If anything the test shows how much of an integral part main memory really is. I take a 14% improvement from 25% more bandwidth to mean the CPU spends more than half the time waiting for memory, when paired with too powerful a GPU.

 

[frozentundra123456]

Well, I was referring to the data you made such a big point about in the excel spreadsheet. If you dont think the data is reliable or "real world" why did you use it yourself to prove your point?

 

[know of fence]

If you dont think the data is reliable or "real world" why did you use it yourself to prove your point?

I hate how Eurogamer juxtaposes two core-i3 with a 300 MHz difference, it's an unfair comparison that makes Skylake look better than it deserves. But their data wasn't completely useless, because nobody else compared HW memory scaling and SL next to each other, while also creating a big enough bottleneck for the difference to be noticeable. Going by 14.5% vs 13% the improvement is essentially the same for SL and HW.

So my verdict is, high memory bandwidth and low latency is what CPUs crave - regardless of architecture. Skylake's strength is: it coincides with the big leaps made by DDR4.

This is the eternal struggle of tech reviews, isn't it it? An electric cattle with twice the wattage will boil your water twice as fast, but it will hardly impact the time to prepare tea, if you also have to dip a tea-bag inside the cup for 5 minutes.
Going from 1600 MHz RAM to 3200 DDR4 is a real 100% improvement and a big deal in a technical sense, but the "real world" impact may be small.

 

[nerp]

IMHO the difference between 74 or 79 fps or 99 or 100fps means the memory difference is negligible for non OC systems.

 

[Headfoot]

But how much sense does it make to combine an i3 with premium RAM, when you can just get an i5 and use the cheapest RAM?

Plenty when the faster RAM is less than $20 more. From newegg, cheapest 2133 DDR4 2x8GB kit w/ heatspreaders = $68. You can get 2666 for $75, and 3000 for $80. It's hardly any more money to move up. I don't know why you'd leave potentially large percentages of FPS on the table to save $12.

http://www.newegg.com/Product/Product.aspx?Item=N82E16820144922

 

[Headfoot]

Nice memory overclock! Uncore?

Still working on that, I've brought it up some but not a lot. My kit was rated for 3200 and the XMP profile put my blck strap at 125 right off the bat which was kinda interesting

 

[Face2Face]

Still working on that, I've brought it up some but not a lot. My kit was rated for 3200 and the XMP profile put my blck strap at 125 right off the bat which was kinda interesting

Yeah same with my 3000MHz kit when I used the XMP profile. I had it working @ 3000MHz for some time, but it refused to boot after awhile. I'm chalking it up to my CPU not liking the 125 Strap, and the PC won't boot @ 3000MHz without XMP, regardless of DIMM voltage. I got my Uncore @ 3GHz. I'm going to take some time so I get focus on getting it closer to the 4GHz number - Stable.

Not used to my below avg 5820K, especially when coming from a Golden 3570K 😀

 

[Termie]

Guys, try DDR4-3200 with manual 16-18-18-40-2T, 1.35V timings on a 100MHz strap. It might work depending on your kit.

 

[Face2Face]

Guys, try DDR4-3200 with manual 16-18-18-40-2T, 1.35V timings on a 100MHz strap. It might work depending on your kit.

I'll give it a shot when I get a chance. Thanks!

 

[Dasa2]

@frozentundra123456
+10% improvement under artificial conditions, mind you, a permanent bottleneck caused by a 1200$ GPU "deep learning" GPU running 1080p resolution on a crap core-i3 (Eurogamer)
+2-3% under sensible conditions and a dual channel high-end setup (TBG/Termie's test)

maybe 2-3% on average as there is a lot of games out there that will be gpu bottlnecked running 120fps+ with a stock 2500k but it depends on the game and section of gameplay 10%+ will show up in parts of some games at normal system settings and can be in the 40-60fps range

overclocking ram can be better value than overclocking the cpu or spending extra on a more expensive one in some cases


http://www.eurogamer.net/articles/d...it-finally-time-to-upgrade-your-core-i5-2500k

http://forums.anandtech.com/showpost.php?p=37837493&postcount=14

This one (dragonage) is artificial due to a low res at 2560x1440 the 290 is bottlnecking the system to ~40fps
but if the gpu was powerfull enough to maintain 60fps+ then the ram would be providing bigger gains that ht or the small cpu oc which was about all this chip could handle

 

[know of fence]

I like your test approach, of varying only one thing at once.
A 80% BW improvement from 1333 to 2400 modules translates to 17.6% more FPS, while a 300 MHz clock bump only yields 6.1%, roughly a third.
So it just might be that 1 GHz OC is about equal to a doubling of bandwidth (from quad channel, or going to fast DDR4), in bottleneck scenarios.

This rule of thumb almost applies to some comparisons out there between Haswell-E and Skylake. But games are all over the place, this rule needs refining. Maybe a doubling of BW is statistically equal to say a 600 MHz OC.
One also could test this by clocking down the CPU until you achieve performance parity.
1. measure FPS with old RAM
2. Install expansive ram (or OC it) ---> new FPS highs
3. Under-clock CPU (with OC'ed RAM) until you reach old RAM FPS levels
4. Compare the differences between clock and bandwidth

 

[AntDX3162]

I had like 100 youtube tabs opened. Memory does matter.

 

[Borealis7]

so from my understanding the conclusion is: ram speed matters with weaker CPUs, and overclocking RAM helps them, while stronger CPUs benefit less from RAM speed?

i was looking at DDR4-3000 sticks for a future i7 build, now i'm not so sure...

 

[Dasa2]

so from my understanding the conclusion is: ram speed matters with weaker CPUs, and overclocking RAM helps them, while stronger CPUs benefit less from RAM speed?

i was looking at DDR4-3000 sticks for a future i7 build, now i'm not so sure...

it will help more in the future when you have upgraded the gpu and games require a more powerful cpu
it just wont do anything till the extra performance is needed
3000 costs what $10-30 extra for 16g?

 

[Deders]

I think the general rule of thumb is that at release prices, 2666MHz was the best bang for buck for Skylake. 3000/3200 is optimal for gaming/general use and anything much faster only helps very specific situations, mostly in benchmarks.

Skylake performance shows more improvement with higher ram speeds than previous architectures.