DDR4 4x4GB Overclocking: Motherboard Challenge
Elite Overclocker Michal “Xtreme Addict” Vobozil examines the issues involved with DDR4 4x4GB overclocking at maximum efficiency levels, comparing and testing Z170 motherboards from four major vendors.
Featuring HyperX Fury DDR4 Memory: Today I bring you another article for both the Extreme OC Community [XOC] and daily OC enthusiasts, this time focusing on 24/7 stability with proper (tight) settings with all memory DIMM slots filled. I hope the information gathered in this article will help users to forget about Intel Extreme Memory Profiles (“XMP”) and unleash the full potential of their systems.
My initial plan was to check 24/7 stability for 2×4 and 2×8 module kits using new Hynix MFR 8 Gbit sticks (which are in fact refreshed 8 Gbit Hynix AFR – more about that in a different article). However, in order to achieve this and show the best possible results, I would need to use several of the best overclocking motherboards around today. Unfortunately, despite being 2nd in the OC world rankings and champion of several OC events I was not fortunate enough to get special versions (non-retail) or samples of exclusive-non-mass-produced boards motherboards (e.g. GIGABYTE Z170X-SOC Force LN2 and ASRock Z170M OC Formula). That’s why I changed my original idea and decided instead to check several high-end Z170 4x slot motherboards in terms of their DDR4 4x4GB overclocking capability.
Filling all the DIMM slots on a motherboard can be problematic. Two things to consider are the motherboard’s tracing design and the performance of the CPU’s Integrated Memory Controller (“IMC”). It’s not a big surprise so see that users are checking CPUs to get higher OC headroom, but few daily users are picking up the CPUs with the strongest IMC. This is mostly reserved for memory binners, enthusiasts and testers.
Today I also want to raise another issue – you’d better sit down in your chair or hold on to something…. you may have to start binning motherboards! From my tests different motherboard samples of the same model gave totally different results, especially when we are talking about Command Rate 1T. Of course I realize that no daily user is going to buy a few motherboards just to find the one which goes a bit higher in memory overclocking. Regardless, it’s always better to extend our perspective and learn new things.
XMP in my opinion should be called the biggest nightmare ever for memory vendors. First let’s recall the origins of XMP, a story which starts with NVIDIA who invented Enhanced Performance Profiles (“EPP”) on their LGA775 chipsets. This was then copied by Intel who took the idea and improved it, thus creating XMP. During the LGA775 era this kind of extra feature should have been making users lives easier, but in truth it mostly didn’t work. Looking back at that time we only had roughly 5-10 memory timing options in the BIOS. It was also a beautiful time because DRAM manufacturers were more interested in fighting for lower latencies, than for high memory frequencies.
Later we found several companies competing in the never-ending “MHz race”, bumping frequencies to ridiculously high values, then charging retail prices literally the equivalent of sacks of gold. At the same time these kits were getting bad reviews from tech portals and testers who revealed that memory speed is not relevant in terms of system efficiency. This is all thanks to the amazing XMP technology, where values were as loose as a clowns pocket, showing only 0.5-1% gain in efficiency compared to standard, high-end DDR3-2800C12 kits with stock JEDEC DDR3-1600C9 specifications.
But why I am not surprised? When you look at default XMP values when pushing your memory, we can see some crazy loose values. For instance TRFC – instead of 128 (still I am talking about DDR3 era) we have 480 for instance, that’s four times looser. During the DDR3 era I often found myself laughing my ass off (on several occasions alongside well-known HWBOT moderator, websmile) when looking at Nick Shih’s memory stability tests, proudly announcing the breaking of another “MHz prime stable” record, all in an attempt to prove that ASRock OC Formula boards were real beasts.
It would all be great, amazing and spectacular but timing values like the monstrous 700 TRFC (which “boosts” memory efficiency to the level of Cedarmill… i.e. just empty MHz) just made the whole situation look so sad. Also note that leaving “AUTO” options on most motherboards leads to the same effect as using an XMP profile. That’s why the real challenge is to find the right balance between low latencies and high MHz frequency in order to achieve proper, balanced efficiency.
With newer architectures IMC efficiency is absolutely crucial to getting a truly fast system. With a properly tweaked memory system on Haswell, we could gain 5-10% more efficiency in many applications. Skylake goes beyond that, averaging around 10-15%. I think it’s well worth making an effort to spend time finding the right combination of latency and efficiency. And please remember, XMP is the devil’s tool!
DDR4 4x4GB DIMM memories are, at the same settings faster, than 2x DIMM. This is mostly because we get additional channel interleaving. In short, it is a design made to compensate for the relatively slow speed of DRAM memory, spreading memory addresses evenly across memory banks. Note: it does require higher stability from our DRAM modules, making it a harder job for the board and the CPU’s IMC. We have to be aware that using 4 modules will always offer lower potential clocks than 2x modules. Life is brutal.
Motherboard PCB Layout design
There are basically two methodologies for creating DIMM slots traces on a motherboard. Unofficially I can add that one vendor tried to create a special PCB with two sets of traces so power-users would be able to just select a particular trace path in the BIOS thus offering the ultimate solution for 2x DIMM or4 x DIMM overclocking. Sadly it seems that the technical obstacle was in fact the switching mechanism. Apparently PCIe control plane switches (for instance, those popular with motherboards that use a PLX chip) are not compatible with that solution. So far no one has created a dedicated controller for this.
Memory Timings on Z170 ‘Skylake’ for DDR4 4x4GB 24/7 Stability
Note: I will be using the memory timing naming scheme that is used in the ASRock Timing Configuration tool. This is one of the most popular utilities among enthusiasts and one that also works on all motherboards from all vendors. I won’t describe and explain all the values, I just want to explore a few key settings in the case of DDR4 4x4GB overclocking (which also applies to 4x8GB) for full stability. In fact most users will find it easier just to copy my settings which I am showing in the screenshots, but I also want to highlight the most important ones. Please note that in all motherboards there is often a different order of timings in the BIOS, so please be careful to check the exact name string in the BIOS.
The tricky one is tWR (especially on ASUS) because it’s linked with tWR-PRE timing in BIOS. To achieve a 10-12 value we need to set tWR-PRE at 23, though tWR influences stability a lot. In fact I would recommend leaving that timing on AUTO. The following tertiaries have to be changed together; tRDWR_sg, tRDWR_dg, tRDWR_dr and tRDWR_dd – with the exception of the last one (tRDWR_dd) which should be +1, setting it at the same number as others might hurt stability or the system may refuse to POST.
Regarding tWRRD_sg and tWRRD_dg, we have to remember we need higher values when we fill all the DIMMs compared to when just using two modules. Especially tWRRD_dg demands values higher at about +8. The last ones in the ASRock Timing Configuration Tool , tWRWR_dr and tWRWR_dd respectively can be clocked twice tighter with only two sticks, but in reality they barely affect efficiency.
The most difficult to control and apply are RTLs and IOLs latency. Depending on the board, sometimes we can’t touch either of them. It depends also on the memory divider, or you can try applying settings by intervals – firstly tweaking RTL/IOL initial value, IOL offset and so on. Still, values sometimes differ from each boot (unless we set-lock everything manually but as I said, often it’s just not possible). In order to play with it sometimes we have to turn the motherboard on and off, play with Memory Fast Boot settings or just pray to the OC Gods that this time the board will boot with the desired value and not lock up at 55 code and POST 3 times in a row.
HyperX FURY DDR4 4x4GB (HX426C15FBK4/16 DDR4)
Just as a reminder – I used four better overclocking DDR4 sticks in my last article. Modules that use HYNIX AFR ICs and are well known to memory vendors. During the Skylake platform launch, those were the fastest available DRAM ICs for the Z170 platform. Comparing to older, highly popular MFR modules among Intel X99 platform users, AFR ICs are capable of achieving higher frequencies with slightly looser timings (especially tRCD/tRP), though as we know Skylake IMC has joined tRCD/tRP timing, so MFR were gradually losing their advantage in terms of lower tRCD value over tRP.
In theory all vendors were well prepared for AFR. This thanks in part to numerous events featuring HyperX AFR memories (HyperX OC Takeover [HOT] 2015 Season 3 Final with MSI) and IDF 2015 (Intel Developer Forum) where the ASRock crew (including Splave and l0ud_sil3nc3) broke the XTU quad-core Global First Place ranking.
Test Methodology and Configuration
- Intel i7 6700K CPU & Intel BOX cooler (Find at Amazon)
- HyperX FURY DDR4 4x4GB (HX426C15FBK4/16 DDR4) (Find at Amazon)
- ASRock Z170 OC Formula (Find at Amazon)
- ASUS Maximux VIII Extreme (Find at Amazon)
- MSI Z170A XPOWER Gaming Titanium Edition (Find at Amazon)
- GIGABYTE Z170X-SOC Force (Find at Amazon)
- SEASONIC SS1200-XP Platinum
- HyperX 3K 120 GB (Find at Amazon)
- Windows 7 64-bit Operating System
- HCI MEMTEST
The main stress was put on CAS Latency 12 and Command Rate 1T, in order to achieve the highest efficiency. On each board DRAM Voltage was set at 1.65V. We have to be aware that 1733 MHz (DDR4-3466) was the highest stock bin of AFR memory from HyperX with XMP (so a super loose profile), but comparing to Samsung based memory, frequencies are pretty low (it’s not hard to find 4000+ Samsung based kits). AFR on air doesn’t run high frequency for 24/7 use. Forcing boards to use 1T with 4×4 configuration with tight timings guarantees higher results with 2T and a higher CAS value, but it will hurt overall performance.
The Core i7 6700K processor was clocked at 4GHz with a 4GHz cache and HyperThreading (HT) turned off. 4x4GB results in 16GB of DRAM memory, which is perfect for 4 instances of HCI MEMTEST, each testing 3,500MB using one logical core of the CPU. By those settings I achieved 100% load on the CPU and CPU’s IMC while using almost all available RAM in the operating system. BLCK was set at 100MHz and I was only operating on memory dividers (no one is messing with BCLK clock when we have unlocked CPU multiplier).
DDR4 4x4GB Overclocking Results
I decided to test a high-end motherboards from each of the four main motherboard vendors. In most cases you will find that all motherboard models within a specific platform series from one vendor are pretty similar in terms of the overall design. One main reason for this is that it also makes the BIOS teams’ life easier. That is why we can expect test results on a Maximus VIII Extreme to be pretty similar to other models, like the Maximus VIII Hero or Gene for example.
ASUS Maximus VIII Extreme: DDR4 4x4GB Overclocking
Overall on ASUS I recommend using tWR value AUTO for beginners. Sadly memory profiles in BIOS 4x4GB are not prepared for 24/7 action although they are compatible with old MFR, not AFR (though 2x4GB AFR profiles work fine). Overclocking 4x4GB on an ASUS M8E (proper sample) was a piece of cake, though on AUTO the board boots with really loose IOLs. What is important here is that tightening IOL is child’s play – we just have to set IOL Offset from 15 (stock) to 21 to get tighter stable values.
MSI Z170A XPOWER GAMING Titanium Edition: DDR4 4x4GB Overclocking
Regarding AFR HyperX clocking – I can feel the influence of the HOT Season 3 Final and BIOS tuning performed by in-house overclockers Pepinorang and Toppc. This board was heavily tuned for AFR efficiency and overclocking. This most likely explains the fact that it’s the only board which could run MEMTEST stable with 12-17-17-28 1T timings at a DDR4-3200 divider. Sadly, the MSI DDR4-3333 divider is a bit “broken” and the board boots with super loose RTLs for Channel B and there is no possibility to tighten it up manually, all options either don’t work or don’t have any effect.
There is no such issues when we are using other multipliers (especially even, DDR4-3000, DDR4-3200), the board sets nice tight RTLs and IOLs. Overall if the DDR4-3333 divider wasn’t bugged, MSI would win this round in terms of this 24/7 4xDIMM stability overclock challenge. I tested 3x samples of the MSI motherboard and all performed similarly.
ASROCK Z170 OC FORMULA: DDR4 4x4GB Overclocking
N Difference setting -4, RFR Delay 14 (CHA), RFR Delay 14 (CHB) and after this we can try to tighten IOL Offset CHA and CHB at 21. Than we should find the RTL Initial Value which boots, and after that finally with those settings we will get pretty nice RTLs/IOLs. We should try to lock them manually, firstly start from the IOL putting values directly in BIOS and then check if board boots, if not – cry into your pillow, if yes – trying to do the same with RTL. If it fails booting, crying is probably the only solution.
GIGABYTE Z170X-SOC Force: DDR4 4x4GB Overclocking
Honestly I have to admit the Skylake Z170 platform is not GIGABYTE’s best moment, especially when we are used to getting decent or at times amazing boards (like Z97X-SOC FORCE or X99 CHAMPION). In the case of this particular AFR test – it was a total incompatibility issue. For a test I put 4x4GB Samsung D-die and I didn’t have problems booting at DDR4-3000. This board hates HYNIX, even in 2x4GB clocking. Nothing more to add or comment.
Final Thoughts on DDR4 4x4GB Overclocking
I have tried to included lot of detailed information in this article, combining a memory setting guide with OC tests of four popular motherboards. The results present us with a lot of interesting facts. Let’s look at MSI’s case – a strong connection with Kingston HyperX and full cooperation to make the HOT Final as good as it could be results in a very mature, fine-tuned BIOS for HYNIX AFR memories. This is one of the facts why extreme overclocking is important and how it affects daily enthusiasts – where there is a demand for the best possible performance/result, more engineers are involved into making the product flawless. Please note that for the test I used the newest BIOS, and indeed it included the implementation of all HOT BIOS tweaks.
The example of the Maximus VIII Extreme proves that the ASUS ROG department is still in top notch form. However, during testing I found real proof (because so far almost no one believed me…) that we really do need to start binning motherboards!
ASROCK’s example shows us that sometimes we have to choose between balance, or max performance in a specific environment. GIGABYTE however is living proof that even big giants from time to time get stuck at in impasse – like Gulliver trapped on the beach.
(Disclaimer: Michal Vobozil is part of the HWBOT Pro OC Program, a cooperative platform for the promotion of overclocking, its community, and its partners.)