Skylake Overclocking Guide [With Statistics]
Skylake Overclocking Guide
Welcome to the Skylake Overclocking Thread. Because a lot of people found my Haswell Overclocking Guide to be a useful source of information, I have decided to try my hand at a Skylake guide as well. For those new to Overlock.net, know that you can open an image in a new tab to view it in its actual size.
Obligatory Disclaimer (Click to show)
I am not responsible if your CPU blows up, your motherboard blows up, your computer blows up, your house blows up, or if your life blows up due to my guide.
My recommendations on what parameters are "safe" are my opinions. I do not have many chips I can subject to high voltages and loads to figure out what is truly safe or unsafe.
What is Skylake?
Skylake is the codename of the microarchitecture launched at August 2015 to succeed Broadwell. The unlocked desktop parts were the first to come out. They are the i5-6600k and the i7-6700k.
How are the 6600k and the 6700k different?
It's the same pattern as in previous generations.
Base Clock: 3.5GHz
Max Turbo: 3.9GHz
L3 Cache: 6MB
Box Cost: $243
Base Clock: 4.0GHz
Max Turbo: 4.2GHz
L3 Cache: 8MB
Box Cost: $350
In most cases the L3 cache difference is immaterial. If you want hyperthreading, the choice is obvious but bear in mind it makes the CPU hotter in testing. Will the i7 part overclock better? That remains to be seen. However, if my Haswell Overclock Chart is any indication, the difference will likely be small. If that 100MHz matters that much to you, you should've bought a binned chip in the first place.
- No stock heatsink.
- z170 chipset.
- Socket 1151. Past coolers compatible with Sandy Bridge, Haswell, etc still work.
- Neither the CPU nor the socket are backwards compatible.
- Removal of Fully Integrated Voltage Regulator, or FIVR introduced at Haswell due to issues at low power applications.
- Lack of FIVR may mean lower temperatures (although that's just one factor among many) and could mean better response under extremely low temperatures.
- According to Sin from OCN and Raja from Asus, changes in architecture requires less robust power-handling components. Because of this, boards tend to have a lower power phase count.
- DDR4 is standard. DDR3 and DDR3L support very rare, and no motherboard supports both DDR3 and DDR4. Still dual channel memory controller.
- It is believed that the Integrated Memory Controller in Skylake is very strong, allowing very high speeds without compromises.
- A smaller die, thinner PCB, and smaller process than Haswell/Devil's Canyon.
- Generally acceptable core voltage has increased.
- Expect overclocks to be similar to that of DC. (In this guide I will refer Haswell as Haswell and Haswell Refresh as DC.)
- Base clock can be changed in a very fine-grain way. Instead of "straps" jumping for 100 to 125 MHz etc, now you can increase it 1 MHz at a time. It's also possible to deviate from the base clock more without instability. Changing the base clock here has no effect on the DMI. However, it still affects ram and PCIE. In total, this makes BLCK changes a more useful thing to try.
- In a similar spirit, ram overclocking is less granular. What used to be 200/266 MHz steps are now 100/133 MHz steps.
- Power usage is similar enough to DC/Haswell. Temperatures similar to DC.
- Graphics updated to HD 530.
What about the "x20 PCIE lanes" stuff I've been hearing about?
- The Direct Memory Interface has been upgraded to version 3. This bumps up the spec to 8.0 gigatransfers per second.
- There are "20 lanes of PCIE 3.0". 16 are direct lanes to the CPU. If you want more you need to go to the enthusiast platform.
- The remaining 4 lanes are handled by the DMI. In the past you had 2 GB/s cap, now you have 4 GB/s cap. There's still overhead, so in reality you'll get maybe a maximum of 3.5 GB/s. You can now run a single Intel 750 Series ssd without bottlenecks. DMI PCIE lanes have a higher latency than direct CPU lanes but the difference should be minor. You can now RAID PCIE ssds through Intel Rapid Storage Technology, and it can be a mix of NVMe and ACHI drives, although that is not recommended. Procedure remains tricky.
- The DMI upgrade also means you will be seeing a lot of new USB 3.1 ports, with a chance of seeing USB Type-C ports. Of course, if all of those are running at full blast, expect your PCIE SSD speeds to suffer a bit.
- Usually in motherboards, the top PCIE lane is x16 PCIE 3.0 lane, second is the same, and the third full PCIE lane is for the SSD. Since Skylake CPUs only have 16 CPU PCIE lanes, putting a second GPU in means both GPUs run at x8.
(Credits: PCPerspective at www.PCPer.com)
Haswell overclocked to ~4.5GHz and DC maybe ~4.6-4.7Ghz. It's far too early to say what the average Skylake overclock will be. Clock for clock Skylake has an extra large improvement in x264, of about 28% over Snady Bridge. The chart might be using an old version of x264 though. Other sources report 35% gain. In Cinebench however, that lead shrunk to 11% over Haswell and 25% over Sandy Bridge. Broadwell isn't often discussed due to low TDP, probably low overclocking, and the emergence of Skylake.
Quick Word about Temperatures and Delidding:
Here are some things to think about before the instructions are read:
- Coolermaster 212 Hyper Evo -> Noctua D14 -> x61 Kraken -> Custom Loop
- Do you want to delid?
- Ripjaws 4 were out during the x99 times, which is pre-Skylake. Both Ripjaws 4 and Ripjaws 5 are DDR4 modules, but it is recommended to use the latter. There have been some anecdotal reports of Ripjaws 4 requiring more voltage to be stable on the z170 platform.
- Unlike Haswell, LLC affects core voltage like the pre-Haswell times, and adaptive voltage mode is no longer dangerous under heavy synthetic loads like Prime95.
- On a similar note, input voltage is no longer a setting. Same goes for cache voltage.
- C-states decrease the voltage and in turn power usage during idle. This has a marginal effect on SSD performance.
- The recommended utility for looking at your stats is HWInfo available here.
- Terminology Check:
Uncore = Cache Ratio = Ring Bus (Not technically 100% true, but when people say these things that's what they mean.)
BLCK = Base Clock
100 (Base Clock) x 45 (Multiplier) = 4500MHz or 4.5GHz
Again, changing the base clock affects PCIE and ram speeds. You will have to readjust the ram setting accordingly if you change the base clock.
Skylake is still a new launch. Keep track of BIOS updates. Things may be improved as time passes.
- If you Bsod, you can look at some details from the crash log. BluescreenView will pull up the information for you. If you want, you can download it here.
- VID is the voltage the processor requests. Generally it is not a useful reading in HWinfo. Vcore when read in real-time in a tool like HWinfo is a measurement of the voltage actually given. When you put in 1.3v into core voltage in the BIOS, maybe only 1.25v is given to the cores under load. This discrepancy is called Vdroop. To counteract that you can simply raise the voltage you entered or you can use Load Line Calibration or LLC. This setting impacts the real-time Vcore reading and increases it. Voltage delivered can have very quick drops, so quickly that specialized gear is required to detect it. LLC helps counteract that.
- Can't find Vcore reading in HWInfo? Don't know where F-clock (Fclk) is on HWinfo?
- Whenever possible, change as few settings at a time. If you crash with core, cache, ram, and GPU overclocks enabled, it's hard to tell which overclock(s) caused the crash.
- Increase multiplier by 1 each time.
- The voltage required to stabilize the next multiplier increases each time.
- Write down the settings you've tried for better organization.
Skylake comes with far more freedom to tweak than Haswell, but this also means more complexity when you are in pursuit of the absolute best settings possible.
Blck/Fclk tweaking time!
Blck, or "base clock", affects multiple things including: Fclk, core frequency, cache frequency, and ram frequency. Let's look at each of these things.
Fclk (Optionally with Base Clock Changes):
Fclk is a setting that has to do with the way the GPU contacts the CPU. The default setting was supposed to be 1000MHz, but due to some complications, z170 boards end up having 800MHz as the default setting for the earlier versions of their UEFIs. The ability to adjust the Fclk was added with later BIOS updates. Fclk is a GPU-oriented setting, so CPU benchmarks won't notice a difference. The difference between various Fclk settings is relatively small. It varies depending on the configuration (especially from GPU to GPU), but for a 980ti the difference is within the margin of error. Open the spoiler here to look at Anandtech's results:
Fclk Benchmarks (Click to show)
800MHz was the original default, with 1000MHz supposedly being the default, but it's possible to overclock this further. From what I can gather, the performance gains of overclocking the Fclk is really quite small, even on graphical benchmarks like Unigine Valley. If the difference in performance past 1GHz Fclk doesn't even show under some graphical benchmarks, the difference is really quite small. However, if you have a GTX 770, this section is for you.
The exact settings will vary because motherboard vendors love assigning different names to the same thing. The instructions below are meant for Asus z170 Hero boards but it should be similar enough to whatever you have to make sense.
Fclk Overclocking Instructions (Click to show)
There should be a setting to adjust the Fclk directly in your BIOS, allowing you to set the Fclk to 400MHz, 800MHz, or 1000MHz.
For example, in the Asus Hero z170 UEFI, under "Tweaker's Paradise", there is an option called "FCLK Frequency for Early Power On". With all motherboards there is typically an option for 1000MHz. What this setting actually does is set the Fclk multiplier to 10, and if the base clock is 100, 100 x 10 = 1000MHz Fclk. In HWinfo, under "System Agent Clock" (refer to Overclock Preparation spoiler), it should now read 1000MHz.
So, if you choose to only overclock the Fclk through the dedicated Fclk setting, make sure your base clock makes sense. If you have it set to 1000MHz and you forget about it and you go back to changing the bclk to overclock your core clock, you won't understand why you won't POST at 170 bclk. The answer is that your Fclk has been overclocked to an insane value.
So let me state it again: Your final Fclk frequency is affected by both your Bclk and the setting you've chosen in the dedicated Fclk setting.
Fclk setting at 1000MHz (Fclk multiplier = 10)
Bclk set to 100MHz
10 x 100 = 1000
Fclk is 1000MHz
Fclk setting at 800MHz (Fclk multiplier = 8)
Bclk set to 110MHz
8 x 110 = 880
Fclk is at 880MHz
If you didn't pick a Fclk setting and you left it at auto in the dedicated Fclk menu, then your motherboard will try to adjust the Fclk multiplier so that you will not crash.
Fine tuning Core Clock with Base Clock:
Base clock x Multiplier = frequency in MHz
Recall that multipliers can only be whole numbers. If we only tweak the multiplier, we can only do 4.5, 4.6, 4.7GHz etc. What if I can do 4.5GHz but I cannot do 4.6GHz? Maybe I can stabilize 4.55GHz. To get 4.55GHz we have to change the base clock. The base clock can contain decimals (like 100.1MHz, etc).
Base Clock Balancing Act:
If you want to adjust the Fclk and ram settings on top of fine tuning core and cache clocks with base clock changes, you have to ensure that:
- Bclk that is not too far from 100 to cause instability.
- Bclk when combined with a core multiplier, gives you the absolute highest core frequency that is stable. (Highest priority, unless using GTX 770.)
- Bclk when combined with a Fclk multiplier (4, 8, or 10) has to result in an overclocked, but stable Fclk.
- Bclk when combined with a right memory divider, gives you a ram frequency that is overclocked somewhat near its maximum.
If you are going to adjust the base clock but don't care about Fclk overclocking, please make sure your Fclk isn't overclocked when your base clock is changing. You can check this in HWinfo as "System Agent" clockspeed.
Whichever has a higher performance index is generally faster. If two sets are very close, the higher frequency kit wins.
Cache overclocking is the easiest thing to overclock but has the least impact on performance (comparisons listed in the later sections). By now, everything should already be overclocked and stable except for the cache. Simply set the cache to the same frequency as your core and stress again to check if it's stable. If it's not stable, lower the multiplier by one and repeat. Unlike Haswell, cache can generally get to the same frequency as the core. There isn't even a cache voltage to worry about.
Unlike Haswell, Prime v28 and Linpack are no longer much hotter than other tests. They are still the hottest tests around, but it's not quite as ridiculous anymore. For example, Prime v27.9 is similar in temperature to v28.7, As the settings chart notes, I detected temperature fluctuations in Linpack, IBT, and XTU stress even though the load on the CPU still read 100%.
Just like in Haswell, note how XTU stress is cooler than XTU bench, and AIDA64 varies in temperature wildly based on the settings checked. Without a way to loop the test, applications like XTU bench and Cinebench are not viable stress tests. As expected, custom x264 at 16 threads is hotter than the 4 thread setting, and using more memory for Linpack causes a hotter test. With the Haswell temperature chart I had 8gb of ram to use, and for this chart here I had 16gb.
My temperatures are lower than what most people will observe because I am not running hyperthreading and I am also delidded. My case as good airflow. Hyperthreading will cause a chip to be hotter and harder to pass stress tests.
Below is a hierarchy of stress tests, listed in order from hardest to pass to easiest to pass. For time-to-crash data, please visit the 'Miscellaneous Testing' spoiler. Anything that is counted as easy to pass or even easier are not recommended and will not be enough to be entered into the main overclocking settings chart. More details in the charting form spoiler.
Linpack (Max) (From Intel's website, not from OCCT or any other place or XTU.)
P95 28.7 S
Stockfish (Chess, BMI2 version)
Aida64 (Full Suite)
Walk in the Park:
Booting into Windows
x264 is the recommended and the default go-to stress test for this thread. If you feel the need to use a hotter test that is your right, but know that your overclock may be hampered by that choice. You could forego delidding in many cases simply by switching to x264. The downside to this method is that the overclocking process will take longer because we are replacing a very stressful program and a short test duration with a less stressful one and a longer duration.
'Prime95 is not Certified for Haswell/Skylake/Insert Nonsense Here'
This was a myth that was perpetuated by some Youtubers. For Skylake it seems this has died down. But know that Prime95 will not eat your CPU and spit out the remains.
Below among the list of stress test download links there is a link for "Linpack Package". I have taken Linpack and added Linx GUI to it. You can now use Linpack as Intel originally intended or run the GUI to easily change the test settings.
Stress Test Download Links
I have redone benchmarks to test the performance difference between a high and low cache frequency.
As you can see, a decrease of 100MHz in core clock has a larger impact on performance than a 1,000MHz decrease in cache frequency. Therefore, my position on cache frequency remains unchanged: It is a secondary setting that you should only overclock and worry about once everything else is done.
If you'd like to see Haswell cache frequency testing, open this spoiler.Haswell Cache Testing (Click to show)
And here are the most recent tests for cache frequency that I have done:
The 4.2 vs 3.4 is the cache setting. The core multiplier for this test was x45.
Testing methodoloy in this test is much more well documented by me.
Chess: Houdini 3, 9mb hash, starting position, 5 minutes.
BF3 Multiplayer: 64 player server in a closed map (Canals). Regular gameplay for entire round.
BF3 Campaign: Second misson, following scripted NPC movement.
Enemy Territory: 30 vs 30, Fueldump.
Runescape: GE, World 3. Capturing FPS while stationary. Max detail, non HTML5. x4 AA Bloom enabled. (It seems to use CPU to do AA)
Oblivion 1: Walk out in the wild, through Oblivion gate, to town gate.
Oblivion 2: NPC combat in Imperial City. Several guards/NPC vs Umbra. Spawn 50 player copies and begin combat once Umbra dies.
These were done on tests, as you can see, that vary from CPU benchmarks to CPU reliant games.
- Stop using IBT or Linpack or Prime95. Use our custom x264 test or use something similar like ROG Realbench,
- If your temperatures from core to core vary over 10C, consider a re-paste of your thermal paste.
- What ambient temperatures are we talking about here? Are you sitting in an oven?
- Hyperthreading makes your CPU hotter.
- CPU delidding service is $50 in the United States from SiliconLottery.
- I recommend using D14 or better in terms of cooling.
- Ensure the cooling solution is mounted properly.
My CPU is downclocking under load!
Check your motherboard's power settings. Set them to max.
The Vcore is far higher than what I've set in the BIOS under load!
Your LLC is probably being overly aggressive. If possible, please set it to a lower amount manually.
Here is an IPC test. (Please note that I showed PCPerspective's IPC charge in the general Skylake spoiler and they have a page about IPC gains as well.
The gains in chess were smaller than I anticipated and the gains in gaming were larger than I expected.
Oblivion and Skyrim were tested in a heavily CPU bottlenecked area. For music encoding I turned 2 flac tracks, 1 wav track, and 1 podcast in mp3 to OGG files through Foobar at 70kbps bitrate. Chess was tested with Arena 3.5 GUI and the latest Stockfish beta at the time with Sedat Canbaz's benchmarking position. Also worth noting is that overclocking the ram from DDR4 2133 to DDR4 3000 gave me a 6% FPS increase in Oblivion (CPU bottlenecked game, apparently also ram as well).
Above is a power consumption test. To state the obvious, this is a measurement of the power draw of the entire system as measured by Kill-a-Watt. ROG Realbench and Battlefield 4 use up a lot of the GPU, making the test pointless. Ram stressing is also a part of the power consumption when looking at tests like Prime or Linpack.
The PSU used was an EVGA Supernova P2 1000w. It is platinum rated, but being a 1000w rated power supply, it is not very efficient at such low loads.
Here is a small chart showing the power consumption differences between 1.4v and 1.2v on the Vcore, and idle power draw differences. Some of you may have seen a chart from Overclocking.Guide showing no power consumption differences between different power saving options. That test is relatively pointless because none of the settings changed idle voltage. Running C-states and such and setting the processor to use 1.4v on load is no difference power-draw-from-the-wall-wise compared to simply setting 1.4v manual voltage and putting that under load. While C8 state saved only 35w when compared to an idle 1.4v setting, many people do not turn off their computer or put it on standby when they are not using the computer. 35w-hour saved for 8 hours a day over a year can add up.
But don't turn on C-states out of concern for the longevity of your processor. The processor is obviously idle when C-states would kick in if it's on, so there's really not much difference.
The settings used for this test are modified ones from temperature testing. The ram was clocked back down to 2133. Linpack takes a minute to load up to maximum load, so for the results I have removed a minute from each result.
Notation is as follows:
3:00, 3/5 meaning 5 runs were done, 3 passed an hour, and 2 failed. Of the 3 that failed, the average time to crash was 3 minutes.
Linpack Max - 2:47, 0/5
OCCT S - 3:38, 0/6
OCCT L - 3:52, 0/3
P95 v28.7 8k - 7:00, 0/35
P95 v28.7 4069k - 1:47, 0/35
P95 v27.8 8k - 10:33, 2/5
IBT Max - 47:00, 1/2
x264 16T passes overnight.
x264 16T passes overnight 50% of the time.
OCCT S - 4:34, 0/10
P95 v28.7 8k - 13:41, 2/10
Linpack Maximum - 25:04, 1/4
OCCT S - 2:35, 0/6
OCCT L - 3:14, 0/5
|Average OC||4.68||Median OC||4.70|
|Average Vcore||1.38||Median Vcore||1.38|
*The Vcore values in the graph above and the average Vcore statistic are derived from the Vcore under load column of the overclock settings chart. Non 6600k or 6700k entries are ignored for the purposes of finding average & median OC/Vcore.
Extra Statistics (From the Bottom of the Chart) (Click to show)
This gives you a rough idea of where you stand.
In order to be charted you need to fill out this form:
Modified text in "Obligatory Disclaimer"... Don't want to seem arrogant.
Thank you for checking out my guide!
Feel free to ask questions or provide suggestions!
Please read the guide before asking questions though!
Please do not PM me unless you think I've missed your post!
I love using exclamation marks!
Thanks, I'm powered by compliments.
My Haswell guide did well so I think I'm set up with too high of an expectation. So far it's not going as well as I would've liked... for example the forum glitch that forced me to push out a half-finished thread to begin with, with a large disclaimer saying 'if you post in here, your post is toast!'. I was also thinking about how the Skylake club is already out with a chart of its own. Although granted, when I did the Haswell guide the official thread for that was long open and Sin had his guide as well. I was one of the earlier people to break out the idea of lowering the uncore and ignoring "1:1" ratio with the core, so that helped a lot. There's not much new this time I fear. This combination of Skylake not being available to me but available to many other people and the a Skylake thread already out with a chart will make it harder for my chart to settle in. But I still think my chart is more complete and it should only get better with time. It's hurting me a lot psychologically to see the availability be 8/31 from SiliconLottery, + more time to bin. I guess I should be more focused on doing well at the task at hand. What can I do with my guide that other people have not?
But I digress (extensively).
Right. So I am assuming that your test is a modified version of Angelotti's?
I see it definitely is. I'm going to start testing it right now and some other things. I believe I also saw Raja say that Handbrake is one of the hardest tests for Skylake, and I'll have that ready...
I cant get it stable @4,6ghz with a vcore set to 1.4v... tried different levels of llc and so on and so on.
My old 3770k from 2012 was i running @ 4.7ghz @ same voltage as this cpu.
The temps are not my problem. When im running @ 1.4v im only hitting MAX 79c in one hour Prime95 testing so i could go further if i wanted to.
How much can i increase the vcore before the cpu could take damage? Is it the temperature wich set the limit or is it any maximum vcore?
|Currently Active Users Viewing This Thread: 5 (1 members and 4 guests)|