A mystery for later...

[mackerel]

This is a system with a 6700k in it, to which I just added Kingston HyperX RGB 4000 2x8GB. Mobo is Asus Maximum IX Apex.


This is my old main system with a 6700k in it, which has G.Skill Ripjaws V 3200 2x8GB. Mobo is Asus Maximum VIII Hero.

Both have CPUs running at stock, 4.0 GHz all cores active. I download latest Prime95 29.8b3 and run a custom benchmark for 4096k FFT, HT off. I know Prime95 is bandwidth hungry. Aida reports the 4000 ram as having more bandwidth, as expected. So what of the prime95 results? About 214 iter/s for the 4000 ram, and 244 iter/s for the 3200 ram. Higher is better. The slower ram is a LOT faster at Prime95.

I know one variable which could be contributing in part, that is the 3200 is dual rank. I've tested it on several ram kits, and it was widely reproducible that dual rank per channel gave a nice boost over single rank. Either by dual rank modules, or two single rank modules per channel. Prime95 is a mixed read/write workload, so maybe something in the way they are mixed makes dual rank work a lot better.

Maybe I need to check other variables. Both are at the same level with Win10 patches. I don't think I'm as up to date on bios on the Hero though... so maybe the new security mitigations could play a part in this too. Will need to swap modules I think to more simply rule out other system differences.

 

[EarthDog]

dual rank, secondary and tertiary timings, etc is where this simpleton would start. Subscribed.

 

[mackerel]

There was a ram OC'er on another forum that suggested a particular set of sub-timings affected prime95 performance significantly, I'll have to see if I can find it again (or just ask). I recall he used it to intentionally control prime95 performance.

I did check for newer bios on the Hero, and there isn't any. That's why I'm still on one from April 2018, and it will be missing all the new microcode since then. It was my understanding that Windows can load microcode in software on boot, so it is still unclear if this is a factor or not. I haven't rechecked Apex bios yet. I think I'll focus my testing only on the Apex going forward to take out hardware and software variables.

 

[Johan45]

https://www.overclockers.com/forums...ulti-rank-memory-better-latency-and-bandwidth

 

[mackerel]

That link is going back a bit. I mention 20% in there, I don't even remember the exact number now it was that long ago, but that would have been comparing at equal clock. Here we have 4000 vs 3200, so 25% increase on clock speed alone, yet 12% slower. I need to do further testing but I don't think rank alone can account for that much difference.

 

[EarthDog]

It may not be stable?

 

[mackerel]

The 4000? I didn't write here but as a quick sanity check I did 15 minutes with aida64 stress test, can't recall exact options but memory was selected. I then also did 30 minutes prime95 large. Both passed.

 

[Johan45]

Most who have been benching for any length of time know that in some benchmarks such as GeekBench dual rank can give a real boost (500 points) but 4xxx MHz CL12 will still beat it. 4000 at CL19, on the other hand, can't make up the difference. Maybe if you could stabilize it with tighter timings (CL17?) it would close that gap.

 

[mackerel]

Past testing with Prime95 showed minimal differences due to timing. Within the limits of my testing at the time, at the same speed, loose vs tight timings made hardly any detectable difference. This compares to wider scaling with clock. I'll have to see if I can find those results again.

I believe this is due to the access pattern being largely predictable so it is easy to preload the cache, thus mitigating any timing cost to a large degree.

Edit:
https://www.overclockers.com/forums...g-in-Prime95?p=7907650&viewfull=1#post7907650

I was beating another users 3733 single rank ram in that. There might have been something in the 1st post of that thread but it looks like I used an external image host at the time which is now dead.

 

[Johan45]

That could be true if you're only tweaking the main timings but there others as you mentioned earlier which affect P95 specifically. IMO the timings need to be dropped across the board to have a noticeable effect. Like running at 4000MHz with full timings from a 3600 kit

 

[mackerel]

Fun fact about tRDWR: these timings directly impact AVX. The tighter they are, the hotter AVX is. The looser they are, the cooler AVX is. Those of you that fear AVX, you might be able to use this to your advantage, and make those stress tests easier on yourself. I promise not to judge you.

From: https://linustechtips.com/main/topic/773966-comprehensive-memory-overclocking-guide/

Think I found it, see above. Of course I'm not so interested in stress, but implicitly the more stress there is, the more work is being done...

 

[mackerel]

Had some time to tinker... some random results.


Unfortunately I don't know exactly how I got to this result, but this is almost as fast as the 3200 dual rank in the other system. Back a step, I tried lowering tRDRW by a couple of cycles manually, but it didn't do much. I wonder if extending it might slow things, but shortening it beyond a point isn't helping me further. While I was poking around the bios, I found various presets for B-die. Some required over 2.0v and said I had to enable LN2 mode. Pass for now. I tried one profile, wouldn't boot. Tried another profile, also no boot. Tried the last profile named something like 4000 1.8v. But I already have 4000 at 1.35v, so why try this, oh, the timings were quite a bit lower. It got to Windows before it BSOD. So close... back to bios, I decided to relax CAS by 1 to 14, and it was benchable.


This is the XMP 4000 results for comparison. I actually had difficulty going back to XMP. It looks like because many timings were set manually, only selecting XMP did not reset those back to auto, and it wouldn't boot. I had to reload bios defaults to quickly reset everything.

Now, you might also notice the 4000 1.8v profile was actually running 3866. I suspect the bios is failing to set all required parameters when using the built in profiles. So it was actually a blend of the 4000 1.8v profile and whatever I had selected before that. Going from XMP 4000 to 4000 1.8v profile also didn't boot. When I got in next time, I noticed it didn't change the voltage at all. At this point I decided to give up with the profiles since they're not even consistently loading.

A quick look at Woomack's thread inspired me to just turn up the clock, where he got 4500 on XMP. Could I try that? No, the bios list only goes up to 4266, and that didn't boot. Back to 4000, I'm now trying to drop the primaries...

 

[mackerel]

A quick update for the night. I still haven't extracted the dual rank modules from the other system due to its location and usage. Might be one for the weekend.

Instead, I put the 4000 modules into my 8350k Asrock Z390M-ITX/ac system. It wouldn't boot at XMP 4000, but did with XMP 3600 profile. It gave better results than I was getting at 4000 on the other system. Both 8350k and 6700k run all core 4 GHz and have 8MB cache, so they should be equal. Maybe it is the combination of timings that affects it. I'll have to put the ram back into the 6700k system and try 3600 on that later. There was a bios update for the Asrock, including vague statements about microcode and ram compatibility, so I put it on. Still wouldn't boot at 4000. Comparing the values at 3600 before and after bios update, a couple of the timings were relaxed a cycle, but I'd have to do a more in depth comparison of timings later.