DDR3 RAM-Myths enlightened

titel_2

Extreme overclockers know about the importance of the right memory kit and the correct memory settings. Apart from CPU and GPU, the memory can have a huge impact on benchmark results. Being part of several online communities, I read a lot of stuff every day such as “just buy the cheapest memory kit, because more than 1600 MHz and 1,50 Volt won’t work anyway with your CPU” which is simply not true in several ways. To help you guys find the correct memory kit, I will enlighten some myths in this post.

Myth #1: You will lose the product warranty if you use your kit outside the specifications

A memory stick is quite simple and only contains few different parts. Basically, a PCB (printed circuit board), the memory chips and a SPD-EEPROM (Electrically Erasable Programmable Read-Only Memory). The SPD-chip stores all important information such as memory clock and latencies. Most of the “memory vendors” are actually just “module-vendors”, because they only buy the parts and solder them onto the PCB. Only few manufacturers such as Samsung or Micron actually produce their own memory chips. So technically it’s very common that you can buy modules from different vendors such as Corsair, G.Skill, Kingston or Avexir which all use the same memory chips e.g. from Hynix or Samsung. This leads me to a very interesting point which I want to clarify with two examples:

Kit 1: Avexir Core Series LED DDR3-3000 MHz CL12-14-14-35
On this kit you will find SK Hynix H5TQ4G83MFR memory chips – so-called MFR. According to the datasheet, the ICs are specified at clocks up to 1866 MHz CL13 at a maximum of 1,80 Volt. Avexir sells these kits at a clock of 3000 MHz with a voltage of 1,65 V.

Kit 2: Corsair Dominator Platinum DDR3-2666 MHz C10-12-12-31
This memory kit contains Samsung memory chips K4B2G0846D-HCH9 – so-called HCH9. These chips are specified at 1866 MHz CL13 at up to 1,975 Volt. Corsair sells them at 2666 MHz C10 using 1,65 Volt.
I’m sure you noticed that the module-vendors already overclock the memory-chips. But how is that even possible?
The module-vendors buy a huge amount of memory chips and test how high these ICs can actually clock. That’s the only way to provide high-end and good clocking kits to the market.
So the result is that it doesn’t even matter if a kit is rated at 1,50 or 1,65 Volt. Depending on which memory chips you have on your module, you can often use up to 1,80 V or even more and the sticks won’t die.
In addition, it also doesn’t matter if you overclock the sticks. You can increase the clocks or tighten the latency and it’s simply not possible to check if you overclocked the memory sticks or not. So no worries about the warranty!

Myth #2: A memory voltage of over 1,50 Volt will straight damage the IMC of your Ivy-Bridge, Haswell or Devils-Canyon CPU and/or the memory kit.

With the second Intel Core generation Bloomfield (Socket 1366), Intel moved the IMC (Integrated Memory Controller) from the Northbridge into the CPU itself. Already for Socket 1366 CPUs, Intel limited the maximum DRAM voltage to 1,65 Volt. According to Intel, everything above 1,65 Volt is outside the specs and might lead to a permanent damage of the CPU. For Sandy-Bridge the voltage was lowered to 1,55 Volt and for Haswell and Devils-Canyon to 1,50 Volt.

However, extreme overclockers often use DRAM voltages of over 2,0 Volt – even on air. You have to keep in mind that increasing the core voltage of your CPU is also outside the specifications, but still everybody does it to overclock the processor.

A lot of people out there will tell you that running a benchmark at a higher DRAM voltage is not the same as the daily usage which I agree on. That’s why I tested my setup for 30 days straight with Prime95 and a DRAM-Voltage of 1,75 Volt. If you use your computer on full load for 2 hours per day, this equals a total usage of over one year.

Setup:

  • i7-4770K retail
  • Alpenföhn Gotthard CPU cooler
  • 2 x 4 GB Kingston HyperX Genesis 1600 CL9 1,65 Volt @ 1,75 Volt
  • GIGABYTE Z87X-OC
  • ASUS GTX 780 DC2

During the test I restarted the setup once every day to have a more realistic result. After the 30 days the system still ran fine without any issues. Safe to say that a voltage of 1,65 Volt is fine to use for the latest Haswell CPUs.

 

Myth #3: Ivy-Bridge, Haswell and Devils-Canyon CPUs only support a maximum memory clock of 1600 MHz.

The 4th gen Intel Core i7 CPUs feature two 64-bit memory channels and support up to 4 ranks per channel. According the the Intel datasheets, the latest CPUs support a memory speed of 1333 MHz (CL8 and CL9) and 1600 MHz (CL10 and CL11). However, a lot of people use memory speeds or 2400 MHz or even higher, so how is that even possible?

Intel guarantees a memory speed of 1600 MHz on every CPU, no matter how many you test or which model you use exactly. If you want to use e.g. 2666 MHz, it could be that maybe 5 out of 1000 won’t do it, but the majority will still be able to run memory speeds which are much higher than the guaranteed 1600 MHz.

According to my testing, 2800 MHz worked on every single 4770K without even using high VCCSA or VCCIO voltages. If your CPU has a bad IMC and won’t run high clocks, it will never lead to a broken IMC. The worst case is just an unstable system at the given clocks. Lower the clocks again and you will be fine. You will also not lose the warranty by running higher memory clocks on the IMC.

According to my testing these memory speeds will always work 24/7 stable:

  • Ivy-Bridge:
    • Core i3: 2000 MHz
    • Core i5 & i7 (non K): 2133 MHz
    • Core i5 & i7 (K-Suffix): 2400 MHz
  • Haswell and Devils-Canyon:
    • Core i3: 2000 MHz
    • Core i5 & i7 (non K): 2133 MHz
    • Core i5 & i7 (K-Suffix): 2666 MHz

Myth #4: “Overclocking-Memory” with 2800 MHz and more is the best stuff available on the market.

Early 2014, a lot of module-vendors released very high clocking kits at 2800 MHz or even more. Already during Computex 2013 in Taipei, I had the chance to test very high clocking memory modules at the booth of Avexir. In the end, I received a kit with 2 sticks which was specified at  3200 MHz C12 at 1,65 Volt. In total, Avexir used 16 x MFR ICs (see Myth#1) on two sticks, so 8 per stick. This means that only one side of the PCB has memory chips on it. We call these sticks “single-sided”.

Memory sticks are divided into ranks. Ranks are basically administration units which can be used by the CPU. Like I already wrote in Myth #3, current Intel CPUs feature up to 4 ranks. This means that a system using 4 ranks at the same time will have the best performance. However, you have to have all 4 ranks on 2 memory sticks. Otherwise, you will not be able to use the dual-channel mode which will also lead to a loss in performance. This also means that using 2 memory sticks in dual channel with only 1 rank per stick will also lead to a loss in performance.

Single-sided and single-ranked is not the same! It’s still possible that you only use one side of the PCB but still divide the stick into 2 ranks. However, all single-sided current DDR3 kits on the market are also single-ranked. That’s why we can say that in general single-sided = single-ranked.

Single- / Dual-Ranked in Detail:

Ranks are administration units which are formed out of a set of memory chips. Usually 8 ICs form one rank.

Information stored in RAM cells are just charged capacitors and due to the low capacity it will lose the information after few nano seconds. To keep the information stored it’s necessary to perform a refresh cycle every few nano seconds.

Rank interleaving is a DDR3 feature which allows to perform a refresh operation in one rank and a access operation in the other rank using dual-ranked memory modules. Single-ranked memory modules can only perform one operation at the same time which will lead to a slower performance because you have to wait one cycle until you can access the information stored on your stick.

rank_1

rank_3

 

 

In addition to the issue of single-ranked memory sticks, at the moment there are almost only Hynix MFR ICs left on the DDR3 market (talking about high clocking sticks!). These memory chips can clock quite high but only with very loose memory timings. So even if you use high memory clocks, you will still have a bad performance.

If you take a look at the prices of high clocking DDR3 memory of 2800 MHz and more, you will notice that the prices are insanely high. Currently, the best bang for the buck are 2 x 8 GB kits with 2400 MHz. For normal gaming rigs everything above is too expensive for the gain in performance. Everything below is only slightly cheaper and you will lose performance.

Almost all DDR3 memory chips on the market have a size of 512 MB. This means that most of the 4 GB sticks are single-sided = single-ranked with 8 chips in total. However, all 8 GB sticks are dual-ranked. So the best way to avoid single-sided sticks is to buy 2 x 8 GB kits.

I ran some game benchmarks for you at different memory configurations to show the performance differences of single- and dual-ranked.

If you use a single GPU system, you will see an improvement of the min-FPS, but almost no difference in the average FPS because of the GPU limit. Multi-GPU setups will have a quite big performance gain, because they are often limited by the CPU.

Another interesting point is that a 3770K using fast 2400 C10 memory will beat a 4770K with slow memory like 1600 MHz single-ranked. So if you upgrade your setup from Ivy-Bridge to Haswell, don’t save money on the memory or you will have no benefit in total!

SR = single-ranked

DR = dual-ranked

1080p_single 720p_single1080p_SLI 720p_SLI

 

Myth #5: There are special overclocking or gaming memory kits

Like I already explained in Myth #1, almost all module-vendors use the same memory ICs across all of their kits. The kits only differ in overclockability, stock clocks, looks and stuff included in the delivery. So there is no such thing like overclocking or gaming memory!

 

Myth #6: Sticks with big heatspreaders are running cooler and have a better performance

The heatspreader usually only improves the look of your kit and definitively doesn’t improve the performance. The power consumption of a memory stick is very low and even without a heatspreader the sticks would work without any problems.

 

Myth #7: Low-Voltage memory will significantly lower the power consumption of your rig

After I had published the German version of this article, I received a request to test this myth. The conclusion is that it is not a myth, but in fact only partially true. A lower memory voltage will lower the power consumption of your rig. However, the power consumption is much more affected by the clock and not the voltage.

I used a 3770K clocked at 4000 MHz on a GIGABYTE Z77X-UP7 motherboard in combination with a Corsair Vengeance Pro 2666 C10 kit and different memory configurations to test this myth.

power

Recommendation to buy:

Rigs with low budget:

Crucial Ballistix Sport DIMM Kit 8GB, DDR3-1600, CL9-9-9-24

Gaming rigs:

Any 2 x 8 GB Kit with 2400 MHz. Preferably C10 such as G.Skill TridentX F3-2400C10D-16GTX

 

Questions or other myths you want to know about? Let me know in the comments!

 

258,940 total views, 33 views today

28 comments

  • taryn

    Have you played around with ddr4 at all in the same sense? Does it follow pretty much the exact same ‘rules’ as this ddr3 that you tested?

    What would be the recommended ddr4 set for medium-high budget?

    • der8auer

      I will also publish a DDR4 guide about single/dual rank and other myths. Some stuff is the same like Myth #1 but single and dual ranked is only a small issue because DDR4 works different.

      For DDR4 the best bang for the buck are 4 x 4 GB 2400 MHz C15 kits. Everything above 3000 MHz will usually not help at all.

  • fat4l

    Hi mate.
    im trying to buy 2x8GB ddr3 as I use crossfire now. My current ram is Gskill TridentX 2800MHz CL12 which is I believe single-ranked.
    I m trying to get some extra performance thus buying 16GB of ram but i need them to be double ranked(this is what adds performance right).

    What i can get is
    G.Skill TridentX 16GB (2x8GB) DDR3 PC3-19200C10 2400MHz Dual Channel Kit (F3-2400C10D-16GTX) 130£
    G.Skill TridentX Fan 16GB (2x8GB) DDR3 PC3-21300C11 2666MHz Dual Channel Kit (F3-2666C11D-16GTXD) 155£
    Avexir Blitz 1.1 Original 16GB (2x8GB) DDR3 PC3-19200C10 2400MHz Dual Channel Memory Kit (AVD3U24001008G-2BZ1) 140£
    or these in promotion for just 110£ :
    TeamGroup Xtreem LV 16GB (2x8GB) DDR3 PC3-21300C11 2666MHz Dual Channel Kit (TXD316G2666HC11CDC01) 110£

    My cpu is 4790k 5G, my GC is 295X2, I play in 1080p.

    THx for helP!

    • fat4l

      Forgot to say, im not sure if all of the above ddr3 are double ranked.
      As u can see they are CL10 2400 vs CL11 2666.

    • der8auer

      Hi!
      Yes if you have 2 x 4 GB 2800 C12 it is for sure single ranked. If you upgrade to a dual ranked 2 x 8 GB Kit you will have the dual ranked speed advantage and the advantage of more memory in total.

      All the kits you mentioned are double ranked because they use 8 GB per stick. The TeamGroup 2666C11 has the same speed as the TridentX 2666C11 so 110£ would be a great price. Even cheaper than the 2400C10. Go and get one of the TeamGroup kits :)

      • fat4l

        Okay thanks ! :)
        Im going for them.

        BTW Teamgroup 2666C11 uses Hynix ICs tho. TridentX 2400C10 uses Samsung ICs as far as I know and also they most prolly will run at 2400C9. However 110£ for 2666C10 is unbeatable! :)

  • Dan

    Myth #1 You didn’t provide very much proof of your statements. I’m pretty sure the OEMs do long term accelerated reliability testing on the ICs at various voltage and frequencies to arrive at the specs. Do you have real proof to prove otherwise?

    Myth #5 What about ICs chosen from specific vendor corner lots? What about binned kits? (eg GTX2) Wouldn’t specificity chosen ICs for their “overclockability” mean they are “special overclocking” kits?

    • der8auer

      About Myth #1: Sure they do testing with the ICs. Still they often run the ICs outside the actual specs and the point was to show that it doesn’t matter if you overclock your sticks or not if they are already overclocked. I can take pics of a several sticks with the given specs and the ICs on it if you want :)

      About Myth #5: Of course you can buy memory at different “binning-steps”. However it doesn’t always guarantee the best overclocking. Corsair had a lot of great kits with the early GTX series but corsair also changed the ICs on different GTX kits. Like GTX 8 v5.11 has Hynix ICs on it and is quite bad for overclocking. Whereas GTX2 had Elpida Hyper ICs and was good for low latencies. The myth just says in general that higer stock specs won’t guarantee you better overclocking results which is true.

  • mutantmagnet

    To my understanding a rank represents the how many times an electrical signal has to “circle” the ram module. As a result a single rank performs half the revolutions of dual rank modules.

    But you talk about ranks in a way I’m not familiar with. You asserts that rank is dependent on how many sticks you are using but rank as I understand it doesn’t require that. Am I right to believe that in the test you did that you used a single stick in the single rank test and 2 sticks in the dual rank test?

    • der8auer

      Hi!

      I just added more detail information about how single/dual ranked works in Myth #4. The key to more performance is rank interleaving :) Please let me know if you need any further explanation.

  • Zahi

    Hi,

    I’m using relatively old Corsair dominators in quad channel 4x 2GB CL8 2000Mhz(2 kits of CMG4GX3M2A2000C8)with a 4770k. Am I loosing performance with this setup?

    Thanks

    • der8auer

      Hi Zahi,

      it depends on what kind of GPU(s) you use and what exactly you are doing with your setup.

      Also did you set the memory to 2000C8 or are they just running on auto?

      The benefit of faster memory might not be worth the price tho.

      • fat4l

        Well he can still sell his currect DDRs and buy a faster set, preferably on promotion. Also now, he is running 4 sticks so he automatically loses dual channel performance.
        But yes, it depends what u are doing on ur pc

  • r0llinlacs

    You say single rank ram isn’t the same as dual rank RAM performance wise but you didn’t mention that single rank RAM would be best suited for 4 stick systems if a CPU can only utilize 4 ranks, am I correct?

    4 sticks of single rank would equal 2 sticks of dual rank, so you could theoretically get the same performance as 2 dual rank sticks from 4 single rank sticks and have more memory at the same time?

  • r0llinlacs

    And sorry for double posting, but that would make 4 sticks of dual rank RAM pointless if the CPU can only utilize 4 of the 8 ranks?

  • Richard Summers

    Hello and thank you for educating me in a honest and concise manner on the greater truths about memory. I have the same question regarding quad channel applications, which is better application with all things being equal in terms of the memory stick. 2X8gb or 4X8gb do I lose performance if I run quad, or no. and how is the answer to my question effected by the fact that quad provides double the amount of memory…

    I am a novice on such matters, but I am assuming that since a properly spec’d system I74770K OC’d would stage data in the additional memory which is the fastest storage application of data, then staging the data in any other part of the system, is this correct? I guess what I am trying to learn is.. Does quad applications slow CPU processing down, and if it does, is the offsetting factor of having a storage space of 16gb of memory to stage data in effective/efficent enough to overcome the negative effect of “quad” vs “dual” applications… Once again, thank you in advance for the insightful education you’ve provided me, I am most grateful for your knowledge and time you afforded me.

  • jeff

    Kinda a noob question, but without overclocking is higher rated ram faster than slower?

    such as a DDR3 1333 vs a DDR3 2400 as an extreme spread using auto settings or the only way to get any gain is through actual overclocking?

  • Sergio

    Something about “dual ranking” doesn’t fit recent Passmark’s results (http://www.passmark.com/baselines/top.html)

    In the Passmark results, 1st to 4th performers are “single rank” 16GB (4x4GB due to the X99 quad-channel) i7-5960 (I guess DDR4). Are really “single rank” or DD4 are different?
    The fifth is a i7-3960 “dual rank” 32GB DDR3.
    It looks like Broadwells perform better with 16GB single rank (?) DDR4 (even against other Broadwell “dual rank” 32/64GB DDR4!!!
    Then Ivy Bridge come with dual rank 32GB DD3 configuration.
    Ivy Bridge also overruns Haswell i7-4790 dual rank 32GB.
    Any insight?
    Thank you

  • satish

    Hi There
    I have Asus maximus vii formula motherboard and is dual channel memory with 4 dimm.

    If I run the above motherboard with 4x8gb Gkill trident X 2400mhz, will there be any difference in performance then running with 2x8gb Gskill 2400mhz memory.

    CPU. 4790k.

    Will be using using the PC for gaming and extreme multitasking.

    Lastly can Gskill quad channel kit be used on motherboard with dual channel 4 dimm.

    Thanks.

    • der8auer

      There will be only a difference if the appication uses more than 16 GB RAM. A higher amount of RAM doesn’t change the performance usually. Only if you are limited because of high RAM usage.

      No Quad Channel on that Board. 4790K CPUs only support dual channel.

  • Miro

    I’m looking to put together a new system. MSI 990FXA Gaming board and AMD FX8350. I wanted to pickup some Corsair Vengeance Pro 2400MHz memory (2x8gb). Seeing as the FX can only support 1866, is there any harm to running that memory? (assuming I choose not to OC).

    • der8auer

      No harm in general but some AMD boards have trouble with high memory clocks. I can’t guarantee that you will be able to run 2400 MHz on the memory.

  • Josh

    Man, I’m having a rough go at it. Trying to get the most speed but on the side of stability. Can anyone help out??

    Dominator® Platinum Series — 32GB (4 x 8GB) DDR3 DRAM 2666MHz C12 Memory Kit
    Intel core i5 4670K core voltage 1.176

    Cant seem to get the clocks correct and my system feels like it’s stop and go. Can anyone help me out with some settings?

    • Tradesman

      With a 4670K and 2666 DRAM you may well require a bit of an OC on the CPU, especially with 32GB. You could also try adding a little additional voltage to the MC (memory controller)

  • Igor

    Hi der8auer!
    Excellent article – read with great interest.
    I have a lot of questions:
    You said that you saw the two strips 3200C12 8 gb. In normal mode. What is this company and what model? Only one catalog A-data read about these, but they were never sold.
    What combination of timing and frequency DDR3 is the best?
    2666C10
    2933C11
    3200C12
    Other values? Intermediate?
    for example
    2625C9
    2400C10
    3200C13 … just get them really ….

    I’m grateful to you for the link to the tests or schedules.
    I choose the best memory for the 5775S. Will overclocking.
    Thank you.

Leave a Reply

Your email address will not be published. Required fields are marked *