DDR3+4930K vs DDR4+5820K comparison

[Woomack]

Many of you are wondering if DDR4 is really faster as we see on all press releases of new hardware. Since I got 5820K and DDR4 then I can directly compare it to 4930K and DDR3 at the same clocks.
If you wish me to compare these 2 platforms in other benchmarks then please let me know.

Test specifications:

i7 [email protected] / 6 cores , 4x 4GB G.Skill TridentX DDR3 2400 @ 2400 CL12-12-12-32 1N for direct comparison
i7 [email protected] / 6 cores , 4x 4GB Corsair Vengeance DDR4 2666 @ 2400 CL12-12-12-32 1N for direct comparison

I will start from benchmarks that are showing memory bandwidth and CPU is less affecting test results.
We have to still remember that new platform has nearly twice as fast CPU cache.

AIDA64



AIDA64 is showing that memory read and memory copy bandwidth are much slower on DDR4 than on DDR3. These two values are actually most important for system performance.
All AIDA64 tests are multithreaded.

MaxxMem



MaxxMem Preview is a single threaded benchmark and here we see that memory read is again better on DDR3. Other results are better on DDR4 so at least this is good. Somehow it's not so good in multithreaded environment what is more important for daily work.


Winsat

I have Polish test systems so it may look weird for you Anyway what is most important are numbers on the right.
Winsat is Windows test which you can run typing "winsat mem" in the command prompt. Simply it's how system sees memory performance.
DDR4 is always on the right ( or below when you are using smaller window ).



I've noticed on some other platforms that Winsat likes memory copy and write bandwidth more than read and we can see that here too. DDR4 is achieving 4GB/second better result than DDR3.


HyperPi 32M



HyperPi base mainly on CPU speed but also likes fast memory. It's hard to compare memory speed here but we see that DDR4 is at least not slowing it down.


These results are on similar memory settings or at least at the same main timings. Of course we can run DDR3 at much tighter timings while tested DDR4 can't run much tighter. Looking at Corsair Vengeance LPX thread we see that 2400 CL12 and 2666 CL13 performance is similar. I couldn't make my Samsung based DDR3 to run at CL13/14 so I used 2400 clock at CL12 for all tests.
Most popular DDR3-2666 kits are running at CL11-13-13 what will give better performance especially that available 2666 DDR4 kits are designed to work at CL15/16 like tested Corsair Vengeance.
A lot depends also from memory controller but it's hard to come any closer in a different platform comparison.

 

[Techjesse]

Very informative write up Woomack I don't know how many GPU's you have, but if you could run 3DMark 06 with a R9 AMD GPU and/or CrossFire, it would compare benching power of the new 5820K vs 4930K. As you know 06 is very CPU dependent and the results may be an eye opener to this new tech (5820K) vs older tech (4930K) plus any edge DDR4 will give. Thanks TJ

 

[bob4933]

Any chance you can run geekbench3 on both? If you can clock them the same, that would be fantastic. Thank you so much for sharing this stuff.

 

[Ajay57]

Great job my friend, yeah some benchmarks would be a nice bonus for all to see.

AJ.

 

[Woomack]

I have only one GTX780 that I can't really move right now. 4930K is in my 24/7 PC and simply there woud be too many things to change. If we compare CPU power then performance gain looks about the same as between i7 3770K and i7 4770K. I will add some more benchmarks later today.

 

[Woomack]

Any chance you can run geekbench3 on both? If you can clock them the same, that would be fantastic. Thank you so much for sharing this stuff.

Geekbench3 32bit ( as only 32b is free )

[email protected] + 16GB DDR3 2400 CL12-12-12 -> Single Core: 3714, Multi Core: 22375 http://browser.primatelabs.com/geekbench3/756760
[email protected] + 16GB DDR4 2400 CL12-12-12 -> Single Core: 3973, Multi Core: 23192 http://browser.primatelabs.com/geekbench3/756755

Memory comparison ( DDR3 on the left )



DDR3 again looks better, especially in multi-threaded tests.


Cinebench R15



In Cinebench we see mainly CPU power but I've noticed that after tightening timings on DDR4, scores are going up more than on DDR3.

 

[Ajay57]

If that is the case then better wait for lower CAS Rates like CL9 etc! If they reach that far down, as for DDR3 i have always wanted to test out CL8 with tighter timings, and see how that would preform!

Like this! http://www.corsair.com/en/dominator...-dram-2133mhz-c8-memory-kit-cmd16gx3m4a2133c8

Enjoy AJ.

 

[Woomack]

Memory linked by you is not as good as you think. Specification looks good but there are single sided Micron modules that are almost not overclocking. The same you can find in Vengeance Pro 2133 9-11-11 that in couple of reviews couldn't pass 2200 clock. G.Skill TridentX on 8GB modules 1866 8-9-9 / 2133 9-11-11 / 2400 10-12-12 are better option as you can make something like 2133 8-10-10/9-10-10 in better kits or something like 2400 9-10-10/9-11-11/10-11-11. If you want benching memory then look for a 4x2GB Corsair Dominator GT or RipjawsX 2000-2133 7-10-7/8-9-8/6-9-6 or similar based on PSC. Old kits but should make something like 2400 8-11-8 on X79.

Back to DDR4. I have some weird feeling that current DDR4 IC are about the same as DDR3 but used on a new memory controller. While testing my Corsair kit based on Hynix IC I've noticed that this memory is scalling similar to Hynix MFR on DDR3. Actually Hynix is also making MFR for DDR4.
My test results are like:
- 2400 12-12-12/13-13-13 1.35V while most DDR3 are specified for 2400 11-12-12/10-12-12 or 11-13-13 1.65V
- 2666 13-13-13/14-14-14 1.35V while most DDR3 are specified for 2666 11-13-13/12-13-13 or 12-14-14 1.65V
- 2800 14-14-14/15-15-15 1.35V while most DDR3 are specified for 2800 12-14-14 and more relaxed have no point, 1.65V

Looking at this scalling I can only say that probably at 1.65V DDR4 will be able to make about the same timings as DDR3 from last series. I also doubt we will see anything like 2400 CL8/9. Probably lowest in mass production will be CL12 unless manufacturers change IC but I think they will go for higher denisty/more relaxed timings instead of tight timings.

Actually you made me curious how tight I can set timings at lower clocks and right now I made it run at 2400 CL11-12-12. Will post results later.

 

[Techjesse]

Thanks Woomack, Although I expended more of a performance boost LGA 2013-v3, X99 still Rocks

 

[Woomack]

The main problem here is that my DDR4 are not really faster at higher clocks and comparing to 4930K/DDR3, most benchmarks are showing much higher CPU performance, not memory. Simply it looks like most manufacturers are making all to think that DDR4 are so great while performance in reviews is based on CPU speed and cores amount. 8 core CPUs have higher memory bandwidth in all multithreaded benchmarks while all reviews that we see are on 8 core 5960X. That's also one of reasons why I'm comparing it on 6 core 5820K as it at least has 6 cores like older generation.

What is funny ( or maybe sad , depends how you look at it ), I can use DDR3 profile from Hynix memory and rewrite most sub timings and it will work on DDR4. To make it work I have to use 1.5V while DDR3 kits were specified for 1.60-1.65V but I bet that these DDR3 will also work at 1.50-1.55V at the same settings.

Results at 2400 11-12-12 1.50V and 2666 12-13-13 1.50V will be posted soon in Corsair Vengeance LPX 2666 thread here -> http://www.overclockers.com/forums/...e-LPX-4x4GB-DDR4-2666-CL15-CMK16GX4M4A2666C15

 

[Johan45]

Good job Woomaack. This at least makes for a fairish comparison. Hopefuly the DDR4 can improve slightly over the next few years when it hits mainstream market.

 

[Woomack]

At tight timings it even looks good but I doubt that any manufacturer will release something like 3000 CL11/12. I think we will faster see 3500 CL16.
Also weird thing is that memory dedicated for X99 platform usually makes you to overclock bclk. I could understand it for the highest series but 2800 is already low clock comparing to what we should expect in not far future. Hard to say if it means that soon we will see new, improved chipsets or new platform.

 

[bob4933]

Thats fantastic Woomack. Thank you for posting it!!! 32 bit is fine for me, it correlates directly to some of my programs I've noticed (more is more kinda thing). Looks like I'll wait for ddr4 to cool off a bit.

 

[EarthDog]

Also weird thing is that memory dedicated for X99 platform usually makes you to overclock bclk. I could understand it for the highest series but 2800 is already low clock comparing to what we should expect in not far future. Hard to say if it means that soon we will see new, improved chipsets or new platform.

That is what Dino was seeing as well in his review on that board. I was wondering if it was a board thing, but now at least one ASUS and one MSI does.

As soon as my memory arrives, I will crank on it and see if it flips the BCLK to 125 as well...

 

[Ajay57]

Like i noticed anything over 2400Mhz on X99 will trigger the Bclk up higher to 125 setting. I am glad i got you to look at tighter timings, something i found on X79, like 2133 can be as fast as 2400 with lower Cas and timings its just takes a bit of tweaking.

This is with 4 x 4 = 16gb Quad matched sets thou, but its only talking about small beer in differences after all, but when Benching its the Diff between scores.

AJ.

 

[MetalRacer]

That is what Dino was seeing as well in his review on that board. I was wondering if it was a board thing, but now at least one ASUS and one MSI does.

As soon as my memory arrives, I will crank on it and see if it flips the BCLK to 125 as well...

My RVE doesn't change the BCLK when I load the 2800 or 3000 profiles.
http://hwbot.org/submission/2622742_

 

[Johan45]

My RVE doesn't change the BCLK when I load the 2800 or 3000 profiles.
http://hwbot.org/submission/2622742_

Is that the end of the chip 4.7, with the phase change?

 

[MetalRacer]

Is that the end of the chip 4.7, with the phase change?

God I hope not.

 

[Johan45]

Good , good still tryin it out then. I was hoping there was a bit more in the tank than that!

 

[Ajay57]

Yeah 1.364v seems not too bad for Vcore on that 5960X Metal!

AJ.