Attempting to compare Skylake i5 vs i7

AIDA64 is not showing any differences if you are running it on 6600K or 6700K. It uses cores, not HT. I have about the same results on both CPUs. Have to use something else for comparison. Geekbench is free and is showing memory bandwidth in multithreading environment.

About a week ago I put the latest bios on the Asus with the Skylake bug fix

Click to expand...

I haven't noticed they released any bios with skylake bug fix. I mean there is no info in log so how do you know they actually fixed it ?
I will only say that ASUS had last BIOS releases ready before anyone knew about that skylake bug. It's actually adding support for new low power CPUs. It was mentioned ... in December in 13.xx release for Hero, Gene and some other boards.

Info from BIOS released 28 Dec:
MAXIMUS VIII HERO BIOS 1302
1. Improve system stability
2. Support 6th Gen Intel® Core™ i3-6098P and i5-6402P processors
*Full support of the new CPU requires BIOS version 1402 or later, and VGA driver version 20.19.15.4285 or later

1402 was released 19 Jan.

I'll look up geekbench and will give that a try.

The bios being a Skylake bug fix is an assumption on my part, but it seems to fit. It was my understanding that the fix was to be included in microcode, and sure enough, we see the description for Asus 1402 bios "Update Microcode and improve system compatibility." The bios identifies its date as 7 January, which is the day after Intel said they had found a fix and were communicating it those who need to know. Time scales might be tight, but not impossible.

The MSI motherboards also had a bios release on 22 Jan, identifying as 21 Jan, description including "Updated CPU Microcode."

I suppose to be 100%, we could look up the required microcode version known to have the fix and check these contain that, but it is beyond my immediate requirements.

Geekbench results are in. The trial version only runs in 32-bit. If anyone has the full version, does it perform much differently in 64-bit?

Anyway, I just reset the MSI i7 system configuration to the same as my current Asus test configuration. That is: both i7's are at 4.2 GHz core (HT off), 4.1 GHz cache, 3000 16-18-18-38-390-2T dual channel ram as reported by CPU-Z. So you would expect them to score the same right? For memory performance, the Asus is about 9% faster in single core, and 14% faster in multi-core. So is the Asus doing something to make it faster, or is the MSI not doing something slowing it down?

Note the Asus is currently running lower speed ram than my earlier results, so allowing a bit more for the higher clock it had, it would be in the right ball park difference I was seeing.

More testing with geekbench. In short, Asus at 3200 scores about 5 to 7% more than itself at 3000 (mem clock speed diff 6.6%). Asus 3200 vs MSI 3000 is 23% in multi-core, so less than I'm seeing but still indicative.

I also note my laptop (MSI) which is also Skylake has a new bios released 27 Jan, dated 7 Jan. I managed to check the microcode version before and after, which were 39 and 55 respectively. I looked at the MSI mobo fw 1.7, that said 74. The Asus 1402 reports 6a. Are microcode versions standardised, or can OEMs give them any number they want? Maybe there is another better version to compare.

You've got one ASUS board and two MSI boards -- buying another ASUS board to test would give you a definitive answer.

My edumacated guess is that ASUS is using better base settings optimizations than MSI, risking increased potential instability to get better performance.

If time is money, the test would be worth it.

Replacing a MSI mobo with another Asus did cross my mind, but I'm not buying another Hero just to play with. If I buy a lower end board, there will still be some question if they are as well optimised as the higher end boards.

Actually, in the Asus bios there are options which are suggestive of higher performance at the cost of stability. They actually tell you that! I have declined those and am using what I believe are the safest settings in the Asus. The MSI doesn't have similar offerings.

I'm taking a break from Skylake and have just bought a Broadwell to play with... I've long wondered if that fat L4 cache might offset the need for a good memory system and hope to answer that too. But that's for another thread...

There are negligible (at best) differences between the performance of motherboards. For example, look at my mobo reviews which put them head to head. Outsode of an outlier or two, the difference between them is 1% or less....margin of error type stuff.

The tasks I run depend significantly on the memory subsystem. Most "normal" tasks wont show that effect. Given what I've seen on my 1st Skylake system with higher clock ram, I wanted to try same with Haswell-E quad channel ram. It doesn't seem worth buying right now given Broadwell-E is almost here and I'll maybe try for that, depending on what regular Broadwell does now I got one to play with.

Bottom line is I'm seeing >20% difference with like for like settings between the Asus Maximus VIII Hero and MSI Z170A Gaming Pro as the biggest variable. Same CPU model, same clocks, same ram timings. I have few paths left to explore that doesn't involve taking apart my main daily use desktop.

1, swap the i5 and i7 between the two MSI boards to see if that gives a clue why the new i7 system is a pain to overclock.
2, buy a different mobo for the i7 and see if performance changes.

I would welcome any explanation as to that difference, as it seems excessive. A few %, fine, but this is too much. On the i5 system previously, I found only running 3/4 cores then boosted per-core performance up to my main i7 system levels. Of course that just increased available ram bandwidth per task by 33%. Given the i5 and i7 on MSI both run the same performance (within a few %) it suggests it isn't due to the i7's bigger cache. That leaves the mobo as the most likely suspect.

If you see bigger difference then you simply set something wrong. I have similar results on ASUS Hero and MSI Z170I Gaming on both 6600K and 6700K. Motherboard can read memory profiles wrong or can set different timings under auto settings but it's +/- 1GB/s in memory bandwidth while 3200 memory already should have like 48-50GB/s.
Up to ~3466 there is almost no difference between boards. ASUS has options to improve overclocking at the cost of stability but there are no options improving performance at the cost of stability. Xtreme tweak option can improve performance in some older benchmarks but that's all.

I'd argue against I "set something wrong" and at most there is some remaining optimisation I have not taken advantage of. I would hope defaults shouldn't be that far off, and I've had to reset the latest i7 rig so often during testing I'm rather sick of seeing defaults

My testing so far has been limited to only the "major" timings and clocks. I haven't gone through the pages and pages of minor timings and they have been left to auto. I could add that to my "to do" list... having said that, if buying and fitting another Asus motherboard gets me the desired performance level without having to spend time tinkering and testing endless bios settings, it'll be worth it. I'm more time limited than cash limited.

I'll see if Prime95 built in benchmark might give a good indication. If so then if others could do similarly on Skylake for comparison we might have a better picture of where my systems sit in the grand scheme of things.

Ok, I've had a play with the Prime95 built in benchmark, and it does seem to represent my observations on relative performance. If I take my Asus i7 system as reference, on the MSI systems a single thread is near enough the same the Asus, and 4 thread performance is about 24% down. The two MSI boards gave within a couple % of each other. Note the Asus is currently back at 3200 ram, where the MSI boards are still on 3000 - settings in the table on the previous page.

I would be grateful if anyone with a Skylake system could run the Prime95 28.7 benchmark and send me the ram scaling part for 1024k FFT in the results. It seems to scale well with FFT size so I don't need to compare the rest. Of course, it would help to know the general hardware (CPU, mobo, ram) and settings (CPU clock, ram clock + major timings).

Following is what I get for my Asus system:

Timings for 1024K FFT length (1 cpu, 1 worker): 3.04 ms. Throughput: 329.48 iter/sec.
Timings for 1024K FFT length (2 cpus, 2 workers): 3.04, 3.01 ms. Throughput: 660.31 iter/sec.
Timings for 1024K FFT length (3 cpus, 3 workers): 3.16, 3.15, 3.15 ms. Throughput: 951.81 iter/sec.
Timings for 1024K FFT length (4 cpus, 4 workers): 3.84, 3.71, 3.72, 3.70 ms. Throughput: 1068.93 iter/sec.

And same for the i7 on MSI:

Timings for 1024K FFT length (1 cpu, 1 worker): 2.96 ms. Throughput: 337.45 iter/sec.
Timings for 1024K FFT length (2 cpus, 2 workers): 3.11, 3.09 ms. Throughput: 645.15 iter/sec.
Timings for 1024K FFT length (3 cpus, 3 workers): 3.66, 3.71, 3.70 ms. Throughput: 812.49 iter/sec.
Timings for 1024K FFT length (4 cpus, 4 workers): 4.79, 4.89, 5.02, 5.03 ms. Throughput: 810.91 iter/sec.

I would be particularly interested in systems with ram running at 3000+, core clock less important. I suspect but haven't tested it yet, that ram bandwidth is more important than latency in this application.

Side note: according to this benchmark, with 4 cores loaded the Broadwell i5-5675C at 3.5 GHz with 1600C8 DDR3 is faster than the Skylakes on MSI.

I've swapped the CPUs between the two MSI mobos now.

The current limiting follows the i7 CPU, so it may simply be that isn't a great sample. It works fine at Intel turbo defaults, just needs a bit more tweaking above that. The CPU booted fine 1st and every time with OC CPU and ram settings unlike the 2nd mobo sample.

The boot issues seem to be only on the 2nd mobo. It does it with the i5-6600k too. Can't remember what I said previously, but it only seems to happen with OC ram.

Still tempted to get a different mobo to test with.

What are you finding out with all this testing?

I'm finding maybe buying these MSI boards wasn't the best choice. They lead me to previously incorrectly conclude the cache differences between i5 and i7 were responsible for the performance difference I saw. With the 2nd system it shows it is due to the MSI boards and ram performance, but I have no idea if or how it is possible to get that performance back. I'm looking at 10's of % drop relative to the Asus, this is not insignificant for the amount of cash I've thrown at Intel! Absent anything else to try, it seems easier to get another brand mobo to try than to keep fighting these MSI boards. They might be best recycled to non-OC systems later on.

Are you testing both brands on exactly the same settings or as you said before on different settings ? I just can't see any significant difference in memory bandwidth in AIDA64 or other benchmarks between MSI Z170I Gaming Pro AC and ASUS Maximus VIII Hero. I was checking various BIOS versions and results are really close.
I can only tell you that RING/Cache frequency makes a big difference in memory bandwidth. Even CPU frequency affects that.
You probably know that but when you compare performance then you have to disable all power saving features and set clocks manually as auto values are different on every board.
On MSI I also recommend to unlock maximum power/current limit as at least on my board after overclocking it's causing throttling when is at auto.

In previous testing I didn't see a difference from adjusting ring/cache so haven't focused on it. Does the CPU need to go through the ring to ram, or is it only for cache? For ram limited cases the cache may not be important. I had tried reducing the ram clock/timings on the Asus to match that of the MSI i7 system (at least as far as the major timings) and the Asus was still significantly faster.

If I run "small" tasks that will sit entirely within the L3 cache, the two systems perform within measurement difference of a percent or two. It is only when doing bigger tasks which hit ram hard do I get this big difference.

Since the i7 is now in the better of the two MSI mobos (one doesn't like booting when overclocking both CPU and ram at same time) I will have another look shortly. I suppose to be 100% I should even swap the ram from the Asus to the MSI boards. I really don't want to interrupt my main desktop, hence getting those MSI boards in the first place...

Woomack, if it is not too big an ask, any chance you can do a Prime95 benchmark on your MSI system, with any Skylake preferably at 4.2 GHz core and 3000+ ram?

PARTIAL RESULT!

On a different forum I had help from another MSI owner. We went through various comparative benchmarks and no pattern seemed to be emerging. It was a higher end mobo, but it was performing roughly in line with my Asus system.

After a bit of thinking, I thought stuff it. I extracted the fast ram from the Asus and put it in the MSI. Turned on XMP and straight to Prime95 benchmark. Fast ram back in my main system, I compared results... and they were practically the same! I was seeing individual tests at -2% to +4% variation but that's likely just measurement variation.

Conclusion: Prime finding is more sensitive to ram performance than I ever thought. Latency aside, a 7% difference in ram clock leads to ball park of 25% task performance. I'm now going to go back and see if I can tighten up the latencies and see if that helps. I wasn't previously able to get the ram test stable faster than 3100, and even at 3100 I had occasional errors so backed off further to 3000 which has been error free.

Random thought, I had underclocked the Asus system ram to same speed as on the MSI system, but the performance didn't drop anywhere like that on MSI, so this doesn't fully explain the difference. Possible differences in what the bios picks for auto settings may be a factor.

Do you think it is your software running in the background?

Do you just ask one line questions?

As far as I can tell, nothing else like background processes should be affecting results. As mentioned, non-ram limited results are consistent across all 3 Skylake systems. So right now I'm trying to figure out what influences the ram performance, now I know it isn't the mobo. I have a lot of data, but still not enough data. I find a good indicator of performance to be the 4 core scaling for 1024k FFT in Prime95 benchmark.

Let's normalise testing as follows to be 100% performance, this is my target:
CPU i7-6700k @4200
Mobo: Asus Maximums VIII Hero or MSI Gaming Pro (now proven both same when running G.skill Ripjaws 5 3200 16-16-16-36 ram)
Ring clock 4100

So, only testing on the MSI, I put the Ripjaws 4 rated at 3333 16-18-18-38 back in. I'm only changing ram settings here.

3000 16-18-18-38 = ~78%
3000 16-16-16-36 = ~77%
3333 16-18-18-38 = ~85%
3200 16-18-18-38 = ~83%
3200 16-16-16-36 = ~85%

I haven't looked at the timings beyond the first 4 yet, could one of those have a massive impact?

Note the above is only for benching purposes. I haven't tested for stability.

Cracked it!

I took out the other two modules and put them all into one system. Couldn't boot ram at 3333 with 4 modules (same problem I had previously) but 3000 booted fine. Ran the prime95 bench, and got average 94% of my fast system at 3200 ram. Strangely enough, 3000 is 94% of 3200. Just to make extra sure, I took 2 modules out again, so back to dual channel, and performance dropped once again. I then tested 4 modules once more just to make absolutely sure. It was, within a percent of the last measurement.

But what is the explanation? Is this the rank thing? The only (known) variable here is putting in 2 or 4 identical modules, unless the bios varies some settings depending on the number present.

To recap, the more problematic ram is G.Skill F4-3333C16-4GRRD. This is 4x4GB kit from X99 era. My fast system runs G.Skill F4-3200C16-8GVK 2x8GB updated for Z170, whatever G.Skill did to do that. 2 sticks of the former showed the low performance. 4 sticks boosted it to comparable with the latter kit, assuming near ideal clock scaling.

2 modules, dual channel, 3000 16-18-18-38 = ~75% compared to my fast system as before
4 modules, dual channel, 3000 16-18-18-38 = ~93%

So that's a 23% increase just from having 4 modules fitted not 2. Oh, that's 93/75 as %, not 93-75 in case anyone gets confused.

This is annoying... it means if I want to run at the higher speed I'd need to buy another kit for the other system, and make sure it will be a "fast" kit. For now I will have to re-split the kit as in the short term it is better for me to have both systems running.

If anyone wants me to run any other free benchmarks in these two conditions then let me know.