Performance of 6700K Skylake and high speed DDR4 RAM

[SAAB340 28]

I've got myself a 6700K and also a set of high speed DDR4 RAM. A TeamGroup Xtreem 3866 18-20-20-40 2T XMP 2x4GB kit. (Based on the new Samsung K4A4G085WE-BCPB IC's for the overclockers that are interested). I've been able to Overclock the RAM to DDR4 4133Mhz 18-20-20-40 1T on 1.4V.

 

I thought I'd show what the improved IPC and RAM speeds actually do for FSX performance. 

 

First of all I put it through the FSXmarkCPU test to show the pure IPC related improvements over previous architectures.

 

 

If we disregard the furthest left Skylake result for now, we can see that Skylake gives a small 4% pure IPC improvement over Haswell, a 16% improvement over IvyBridge, a  21% improvement over SandyBridge and a 31% improvement over Lynnfield. The IPC improvement over Haswell is really minor, especially if we consider that Broadwell should be in between as well...

 

 

However, we also have to consider that Skylake uses DDR4 RAM. DDR4 starts at 2133MHz 15-15-15-36 2T. Given that you can't buy DDR4 slower than that, its not realistic comparing Skylake using DDR4 running at 1600MHz 9-9-9-28 1T. 1600MHz 9-9-9-28 1T is a fairly slow DDR3 speed. However you get even slower DDR3 RAM. Especially if you have a kit from early on. Intel actually only officially supports 1600MHz DDR3 RAM on the other listed CPU architectures. Everything else is considered overclocking by Intel. So it's a much better comparison to use the left Skylake result. That gives a much more respectable 11% improvement clock for clock over Haswell, a 23% improvement over IvyBridge, a 28% improvement over SandyBridge and a 39% improvement over Lynnfield.

 

So what does that mean in reality. When your FPS dips low it's normally the CPU that is the limit. So at the same CPU clock, at a time where; Haswell dips to 27FPS,  IvyBridge dips to 24FPS, SandyBridge dips to 23FPS or Lynnfield dips to 22FPS, Skylake would only dip to 30FPS.

 

 

You might say that you can use faster RAM on the other architechtures as well. That is true. But Skylake doesn't stop at 2133MHz, it just starts there. As I said in the beginning, I've been able to use 4133Mhz 18-20-20-40 1T on my 6700K and that is a lot faster than any of the other architectures managed to do with DDR3. 

 

Lynnfield was really tricky to combine a high CPU overclock past 1600MHz on the RAM. The memory controller wasn't strong enough.

 

SandyBridge only have a 2133 memory strap so you can't use memory faster than that. Most CPUs manage that though. That still means that even using cheap DDR4 the 6700K at stock speed performs better than any SandyBridge system overclocked to 5GHz. Did I mention that Skylake overclocks too and can use way faster RAM?

 

IvyBridge allowed for a lot faster RAM. Max memory strap is 3200 i believe. However the memory controller in only 1 out of 4 CPUs is able to manage RAM as fast as 2800MHz.

 

Haswell has an even better memory controller where pretty much all CPUs manage 2800MHz RAM and about half of them manage 3000MHz RAM. 

 

So what about Skylake? It has a very strong memory controller. The word is that pretty much all CPUs manages 4000MHz RAM.

 

So how does this high speed RAM translate to FSX performance? Today I've been busy running a whole lot of FSXmark11 benchmarks at a variety of CPU and RAM speeds/settings.

 

These are the results:

First of all, both CPU and RAM @ Stock speeds.

 

 

CPU @ stock speed. The RAM @ 3000MHz 15-17-17-35 2T. A common XMP setting for DDR4. This is very sweet spot when it comes to pricing of DDR4 at the moment. Above this and the price starts to soar.

 

 

CPU @ stock speed. My RAM @ XMP settings 3866 18-20-20-40 2T 

 

I appreciate that not everyone is comfortable with overclocking. I'd encourage you to learn how to, but it's not for everyone. Some people are not in to CPU overclocking but are happy to go in to BIOS and set XMP. Well, this is performance Skylake has on offer for you if you consider an upgrade.

 

For those of you that are more in to overclocking I'd managed to get my 6700K up to 4.7GHz on 3.75V. My setup can't handle more heat unfortunately. My Skylake rig is a Fractal Design Node 304 mITX chassi using an old but trusty Thermalright Ultra-120 Extreme cooler with a single Noctua 120mm fan in the less than ideal pull configuration. It's not intended for any serious overclocking, but I had to try and see what was possible. I know the serious overclockers will be able to produce even better results. Westman already has. The results I achieved are nothing to be ashamed of though. 

 

With stock RAM

 

 

With 3000MHz 15-17-17-35 2T RAM

 

 

With 3866MHz 18-20-20-40 2T RAM

 

 

With 4133MHz 18-20-20-40 1T RAM

 

 

As you can see, faster RAM does indeed bring more performance. The move to 3000MHz provides the biggest jump in performance, but even faster RAM still helps. You might be saying that the improvement by RAM is so small and doesn't matter. Let me provide you with this final result and you can have a think about it:

 

CPU @ 4.3GHz RAM @ 4133MHz 18-20-20-40 1T

 

Compare 4.3GHz CPU with 4133 18-20-20-40 1T RAM to the previous 4.7GHz CPU with stock RAM. The 4.3GHz CPU one is ever so slightly faster. Just the change in RAM is actually the same as a 400MHz CPU overclock, and that is without actually overclocking the CPU. When you have fast RAM you can still overclock the CPU for even better performance. By going for DDR4 around 3000MHz instead of the slowest 2133MHz RAM it equates to a 200-300MHz CPU overclock. RAM does matter when it comes to FSX. It's one of few programs that can actually make use of faster RAM.

 

There is no magical RAM speed/latency that is needed. FSX is able to make use of faster/lower latency RAM. Just bear in mind though that the price/performance ratio of a high performance RAM kit is appalling. So if money is of concern never pay a lot extra to get faster RAM. But always pay a little extra for a lot faster RAM. DDR3 at 3000MHz is/was very pricy. DDR4 at 3000MHz is a lot cheaper. 

 

 

Thanks to all the people that have contributed to FSXmarkCPU, FSXmark11 and to Westman for providing more details about what RAM overclocks that are typical on different CPUs.

 

[gboz 157]

Excellent work there, thanks for the comparisons.

Did you underclock the TeamGroup Xtreem RAM for the lower speeds?

 

gb.

[westman 257]

Nice work Saab340!!!

G.Skill have mems based on the same samsung chip , later batches of this F4-3200C16D-8GTZ no problem with +3866 mhz confirmed, the price is good.

Not all Mobos do +4000mhz keep in mind that the best MoBo for mem OC is Asrock.

I have a Asus Hero VIII Its ok up to 4000mhz not much more the Asrock OC formula do 4266 with same mems.

Most mobos do 3600mhz a good goal is 3600mhz 16-18-18-38 for not experianced overclockers.

That would be as the Haswell 2400 cl9 recomendations.

 

Got a set of the same Team 3866 and the tridentZ 3200 , have hyper X fury 2666 G.Skill Ripjaws4 3000 these ar based on Hynix MFR they stop at 3200mhz , have also a micron based kit Crucial ballistix not like them.

Almost forget the G.Skill 3400 Ripjaws4 based on the Old Samsung chip ( not as good as the E-chip)

[martin-w 3,235]

Fantastic work! Thanks so much for your efforts SAAB340! Endeavours like this benefit all of us.

 

23% improvement from Ivy Bridge to Skylake sounds great to me. I have components for my Skylake system arriving this morning.

 

I went for Ripjaw V's for ram.

 

3200 MHz. F4-3200C16D-16GVR to be precise.

[SAAB340 28]

Excellent work there, thanks for the comparisons.

Did you underclock the TeamGroup Xtreem RAM for the lower speeds?

gb.

Yes. It's the same RAM used all the time.

Fantastic work! Thanks so much for your efforts SAAB340! Endeavours like this benefit all of us.

 

23% improvement from Ivy Bridge to Skylake sounds great to me. I have components for my Skylake system arriving this morning.

 

I went for Ripjaw V's for ram.

 

3200 MHz. F4-3200C16D-16GVR to be precise.

Thanks. I think Skylake and Ivys in general manage a similar overclocks. Maybe a small advantage to Ivy if you De-lidded. But the RAM should ofset that. Hope you got a good chip.

Happy building!

[SAAB340 28]

Nice work Saab340!!!

G.Skill have mems based on the same samsung chip , later batches of this F4-3200C16D-8GTZ no problem with +3866 mhz confirmed, the price is good.

Not all Mobos do +4000mhz keep in mind that the best MoBo for mem OC is Asrock.

I have a Asus Hero VIII Its ok up to 4000mhz not much more the Asrock OC formula do 4266 with same mems.

Most mobos do 3600mhz a good goal is 3600mhz 16-18-18-38 for not experianced overclockers.

That would be as the Haswell 2400 cl9 recomendations.

Got a set of the same Team 3866 and the tridentZ 3200 , have hyper X fury 2666 G.Skill Ripjaws4 3000 these ar based on Hynix MFR they stop at 3200mhz , have also a micron based kit Crucial ballistix not like them.

Almost forget the G.Skill 3400 Ripjaws4 based on the Old Samsung chip ( not as good as the E-chip)

Looking forward to seeing what you end up squezing out of your system in the end as you learn to master Skylake.

 

It didn't actually cross my mind that the MoBos might limit the RAM 'already' at 3600. Very good thing to be pointed out. I guess we will start to see reasonably priced kits up to that sort of speed. DDR4 certainly have come a long way already even though its still very young and is away to become mainstream.

 

I use the AsRock Fatal1ty Z170 mITX. It looks like AsRock have done a good job with their board designs when it comes to RAM. 4133 is the highest strap I have on offer in the BIOS. Been thinking of making use of the unlocked BCLK to see how far my RAM sticks can go. It won't even POST at any timings tighter than 18-20-20-40 @1.4V. Seen the great results at higher mem voltage though. It's always tempting to bump that Voltage higher when overclocking...

[westman 257]

1.45-1.5v should not harm the memcontroller , but Its unclear if it degreate over time?

The samsung chip is ok for 1.65v

Did a test at 11-18-18-28 1t 3866mhz at stock memcooling prefer to not mention the voltage used.

[Ted Striker 338]

Excellent testing and presentation SAAB340. Thank you for doing this. Being one who is currently on an Ivy Bridge at 4.5 ghz and 2133 mhz memory this provides me a very clear picture of what I can expect from upgrading. Very professionally done.

 

Ted

[SAAB340 28]

Today I've done a further look in to load times and texture loading. I have terabytes of photo scenery myself. Personally I don't use much demanding 3rd party aircrafts or airports. I know that's certainly highly unusual here. So the texture loading improvements that can be gained by using the increased core count in the Intel Extreme line of CPUs is basically irrelevant to 99.99% of people who reads this. But I personally find it an important measurement.

Load times however, the time it takes to load a flight, is still a relevant metric for all of us. Texture loading and load times are very much CPU limited as long as you use an SSD. So they scale linearly with CPU clock speed.

 

When I got my SandyBridgeE system I had a quick look at how much effect RAM had on texture loading and load times. The finding then was that it didn't really matter. Today I had a look with the Skylake system @ 4.5GHz and used 2133 15-15-15-36 2T and 4133 18-20-20-40 1T RAM for the comparison.

Basically, RAM still doesn't matter for texture loading or load times.

(You can't see it without looking in to the decimal points, but by taking the average from 10+ measurements in each test that I've done here, there is a tiny trend that emerges of about half a % improvement by using the much faster RAM. The variance between each of the measurements is still a lot greater than that and half a % is in reality nothing.) 

 

I also tested at 4.3GHz at 4 relevant AffinityMasks so I could compare against the SandyBridgeE measurements I've made in the past. My Skylake build uses the same Intel X-25M G2 SSD for OS/FSX that I used to have as my OS/FSX disk in my SandyBridgeE system before I upgraded a while back. So today I simply moved across my 1TB Samsung 840EVO SSD that has all the scenery needed for the test and got to work.

 

This is how Skylake compares to SandyBridgeE (and very likely to regular SandyBridge as well).

 

When it comes to load times and texture loading I guess I should mention that AffinityMask=84 on a Hyperthreaded system performs the same as AffinityMask=14 without Hyperthreading. And AffinityMask=85 with Hyperthreading performs the same as AffinityMask=15 without Hyperthreading.

 

From these results we can see that the texture loading improvements are fully in line with the FSXmarkCPU improvements while using the same RAM settings for Skylake and SandyBridgeE. In texture loading the 4 core 8 thread Skylake still can't beat the 6 core 12 tread SandyBridgeE, but it's getting close. Especially if we factor in the higher possible overclock for Skylake. But as I said before, this is irrelevant to 99.99% of all of you. I know of photo scenery lovers that have left the 8 core 16 tread HaswellE for Skylake and they are very happy with the result. Pretty much all the Intel-E CPUs are poor Overclockers. It's a lot harder to get 6 or 8 cores that all can achieve high clockspeeds on the same chip. So the overclock and the all important single threaded performance that is what matters most in FSX does suffer.

 

What this also shows though is that Skylake offers significantly improved load times over SandyBridgeE. I thought load times were speedy on the bare install we use with FSXmark11 but I thought I'd just forgotten how fast things load without all the add-ons I usually have. But this test confirms a very significant speedup over SandyBridgeE. I've double checked and I have the same add-ons installed on the same disks as I had with the SandyBridgeE when I tested it previously. I used native SATAIII of the chipset with SandyBridgeE as well. The speedup with Skylake is 60-70%, getting larger the more threads that are used. Is this all due to Skylake or is it the sum of a few large, but gradual speedups through the generations of IvyBridge and Haswell? I can't say. Anyone that has upgraded from those platforms that wants to chime in?

 

What I do know is that I'm really enjoying it. I've shown before how much using SSDs instead of HDDs for scenery speed things up. The upgrade from Lynnfield to SandyBridgeE didn't offer anywhere near this load time speedup. Even compared to using all the 6 hyperthreaded cores on SandyBridgeE, Skylake still cuts out 1/3rd of the load time using its 4 hyperthreaded cores. That is very much noticeable. On a HDD I'd not expect as much of a speedup as the HDD will be the botleneck, but I'd still expect to see a noticeable improvement.

 

Now its time to do more flying and less testing on this Skylake build. I've got a brand new SSD waiting to filled with scenery =). My wife just asked how long I'm planning to spend all my free time in front of my new toy as she claims I've only said about 10 words to her the over the last two days. Don't think that is totally accurate but she has a point. Happy flying!

[gboz 157]

Yes. It's the same RAM used all the time.

 

 

Ok, thanks.

 

I guess your overclocks are with all cores overclocked.

Have you ever done a FSXMarkCPU with just a first core overclock, (hopefully at higher OC than with all cores) and compared?

 

gb.

[Adam Reed 29]

Thank you for taking the time to do these tests for us SAAB, much appreciated. 

[Makinen11 184]

Extremely well documented testing, it gives me a good idea on what memory to buy for my upcoming skylake build as well. Thanks Saab, we appreciate it!

[SAAB340 28]

When something looks too good to be true it normally is. That still applies. Unfortunately there's no massive load time improvement with Skylake. Even though I double checked, I basically messed up. There were several MSE2.0 states missing in the scenery.cfg I found out when I checked it over once more. Therefore, I've had to re-run the texture load and load time test. I've kept the original data in the result image in red and marked it as incorrectly configured as the image will still appear in the original post above, and its too late to edit the text in the post.

 

So here are the updated results:

 

Unfortunately there's nothing exciting in them. The incorrect scenery configuration had only a minor effect on the texture loading but had a larger effect on load times. Basically, the texture loading is still in line with the FSXmarkCPU improvements while using the same RAM settings for Skylake and SandyBridgeE. And now the load times are in line with them as well. Sad but true.

 

There's still a healthy improvement compared to the several generations older SandyBridgeE clock for clock, like for like. But compared to Haswell, the indications are that there won't be much Improvement with Skylake when it comes to texture loading or load times. 

[skysurfer 76]

Hey guys.

 

I am trying to understand all those numbers above, I am still having issues with it.

I am upgrading my PC and so far I have i7-6700k, Nvidia 980. Not sure about motherboard, most likely I will get either ASUS VIII HEro z170 or ASUS Sabretooth z170 m1. From here I need to buy the right RAM. So this is where I am stuck. Will those 2 mobos be OK for i6700k for overclocking?  What will be the good RAM? 

 

Thanks for looking. 

[westman 257]

My simsystem is 6700k and Asus Hero VIII Z170 have differnt setts of mems , i recomend G-Skill Ripjaws or TridentZ 3200-3600mhz be sure it is samsung E-die chips.

The Hero VIII is a very good MoBo for overclocking.

 

Martin W here building a very nice system look witch parts he use , the hero VIII cost a little more but not nessesary for overclocking all asus z170 is basic the same when it come to design have identical memory lines for example , the best Asus board for overclocking whith single Gpu is the Maximus VIII Impact mostly for 2 men slots system ( the only Asus Z170 Mobo with different design for memory lines )