TridentZ 3000C14, memtest vs aida64

[EarthDog]

Its likely too high for the memory clocks you are trying...

 

[mackerel]

I'll aim to give lower voltage a try on the MSI Z170 board tonight, before moving it into the Z370 system I really want to use it in I think I only went as low as 1.20 on both in my testing.

 

[mackerel]

Trying lower voltages... some interesting results...

Still at XMP, I set 17-17-17-auto, 3600, IO 1.05v, SA 1.15v, booted and seemed fine under stress for about a minute.
Reboot, set timings to 16-16-16-auto, booted, stress near instant errors out.
Reboot, set ram to 1.40v, wont even boot at that! More voltage upsets it? I'm thinking, is there something like the Ryzen VTT thing that needs to be half ram voltage?
Reboot, set ram back to auto (1.36v reported), lower SA to 1.10. Booted, stress test running fine after 9 minutes. I think I could call this bench stable, which is far more than I managed to get before.

This is why I never bothered with tweaking ram before. So many variables, and I don't have enough experience or knowledge to make educated guesses beyond looking at typical values or what others might have done. Next stop, 15-15-15...

Edit:

Trying 15-15-15... crash on Windows boot
Lower SA to 1.05, a different sort of fail to boot, single white line (like cursor, not flashing) on screen
Ok, maybe 15 isn't happening. What about next speed up?
Set 3733 17-17-17, SA back to 1.10. No boot.

I went to look at 3600 16-16-16 again, but on trying to save an aida cache benchmark it BSOD. So still unstable after all.

Ok, I'm calling it a day on the MSI Z170 mobo. I got some interesting data at the least. Next stop, the Asrock Z370...

Edit 2:

Some results on Asrock Z370 Pro4:
3000 14-14-14 ok 56m
3000 13-13-13 seems ok
3000 12-12-12 no boot
3000 13-12-12 fails stress quickly
3600 17-17-17 seems ok
3600 16-16-16 had to reduce IO/SA to boot at all, stress testing now. Ok for 10 minutes so I guess I'll have to put SuperPI on it next 3000C13 is in theory lower CL time than 3600C16, but which helps more?

The mobo at least has a 3000 multi that doesn't involve messing with fsb.

 

[Woomack]

Reboot, set ram to 1.40v, wont even boot at that! More voltage upsets it? I'm thinking, is there something like the Ryzen VTT thing that needs to be half ram voltage?

There is but at auto it's always 50% of DIMM voltage.
I guess you need new BIOS for Z370 and Z170 may have some other issues. Z170I Gaming was great for memory OC but full atx not so much.

Run AIDA64, whatever has higher read bandwidth is in general better. Latency is not affecting performance as much as on older platforms. Other way to check how OS sees performance is to run winsat mem from command prompt. It's more like a memory copy result but based on that Windows calculates memory performance.

 

[mackerel]

Just did some SuperPi 32M on it which I need to post in the other thread shortly. The 3000C13 time was about 0.5s faster than 3600C16, but I only did two successful runs of each and took the best. At 3600C16 I had one soft crash so stability isn't perfect.

 

[mackerel]

I need to write things down more often. Some quick notes before I forget again.

3000 13-13-13 seems stable
3600 16-16-16 seems stable (2h+ aida64 passed)

Forgot to note down exact voltages, but I think IO is 1.05 and SA is 1.15

Attempts to get 3700 anything booting have all resulted in failure. Tried slacking timings to 18 and more voltages without luck.

Having some fun against a German on hwbot for the 8350k. We're all hitting 5 GHz and no further, I'm outclassed by ram spec, but have managed to pull some leads in some areas.

 

[Woomack]

You should have much better time on 3600 than 3000. Probably too many sub timings at at auto. Easy way to set it better, I think I already said that before ... set memory at auto/2133, write down timings, set 3600 and use exactly the same sub timings, set tWCL=CL and main timings to 15-15-15 1.5V. Check if all is working, if yes then try 14-14-14 at 1.5V. If you make it work then try 13-13-13 and probably higher memory voltage.
This is the easy way, hard way will take you 2 weeks+ to check all timings and won't be much better anyway.

If you make all right and motherboard let you then you can set 3600+ 12-12-12 or 12-11-12 at high voltage for tests. However there are sometimes issues with stability and for tests you may need to set maxmem value in windows to ~3GB or run 32bit OS. CL13 is much easier to set. Problems are generally starting at CL12. It's more motherboard's fault than the memory.

All these magical overclockers on hwbot run pretty much the same memory profiles made by others. Similar sub timings can be found in asus/asrock OC memory profiles. It's just easy to see that most of the overclockers use already checked patterns and have to idea how to set it on their own.

 

[mackerel]

You had mentioned most of that before, but I haven't had a chance to spend some time on it. A job for the weekend. I was just setting things to try while doing other things elsewhere.

 

[mackerel]

Back to playing with the ram. I saw a comment elsewhere that Coffee Lake IMC likes more relaxed tRAS than previous CPUs. To recap, I'm happy enough that 3600 16-16-16-auto, 1.05 IO, 1.15 SA seems stable, but I never booted at any speed above that. I really relaxed timings, set 3700 and... still no boot. More mem voltage to 1.4... nope. Relax timings even more, still no. Just as I was about to give up, I set 3733 and here I am! I've only done a basic test of 4 minutes aida64 cache+ram and seems ok so far.

3700 would be on 100 strap, and 3733 on 133 strap right? Does that have much possible impact? Now I've made that step, I will aim to try and push more on clock before tightening up timings.

 

[Woomack]

Strap shouldn't affect performance etc.
Check something like 3733 17-17-17-38 1N 1.35-1.40V. I have 4x16GB at these settings for over a week without issues ( PC runs 24/7 for over a week ). Well, on X299 but on Z370 should be easier to set it.

 

[mackerel]

Wasn't questioning performance, but stability. 3700 no boot at all in all the variations I've tried. 3733, 3866 booted. Have you come across the statement of Coffee Lake liking more relaxed tRAS before? This was on Anandtech reviews of two recent high speed kits, where apparently both G.Skill and Corsair have relaxed timings compared to previous. I haven't compared to check exactly what that difference might be.

3866 got to Windows on same settings as 3733, but aida stability test gave error almost instantly. Just got delivery of a SSD so while preparing lunch, I'm cloning the Win7 test install onto that (from HD) which should help speed up further testing.

Edit:

Lunch out of the way, back to testing:
3866 18-22-22-44 1.40v, IO 1.05v, SA 1.15v - failed in Windows
3866 18-22-22-44 1.40v, IO 1.05v, SA 1.17v - failed in Windows
3866 18-22-22-44 1.40v, IO 1.07v, SA 1.15v - ok for 14 minutes in aida64
4000 18-22-22-44 1.40v, IO 1.07v, SA 1.15v - starts, too unstable to do anything in bios
4000 18-22-22-44 1.40v, IO 1.10v, SA 1.15v - failed in Windows
4000 18-22-22-44 1.40v, IO 1.13v, SA 1.15v - failed in Windows
4000 19-22-22-44 1.40v, IO 1.13v, SA 1.15v - failed in Windows
3866 18-22-22-44 1.35v, IO 1.10v, SA 1.15v - ok for 39 minutes in aida64
3866 18-18-18-44 1.35v, IO 1.10v, SA 1.15v - ok for 38 minutes in aida64
3866 17-17-17-44 1.35v, IO 1.10v, SA 1.15v - ok for 11 minutes in aida64
3866 17-17-17-38 1.35v, IO 1.10v, SA 1.15v - ok for ?? minutes in aida64 (got distracted and forgot exactly, over 20 mins)
3866 16-17-17-38 1.35v, IO 1.10v, SA 1.15v - no boot
3866 17-17-17-36 1.35v, IO 1.10v, SA 1.15v - ok for 62 minutes in aida64

I'm now taking a break from pushing the ram and will take some benchmarks in this configuration. Then I'm undecided if I'll either try 4000 again, or start playing with secondary/tertiary timings at 3866.

 

[mackerel]

Status update: I thought I had things dialled in nicely, and started playing a little with tRFC and tREFI. As I prepared for my evening meal I set prime95 on itwith CPU overclock, and things started to get interesting. On suggestion from elsewhere, I was doing additional testing with prime95 set to fft size 512k upwards. That was fine.

For small FFT, it was stable to a tested 4.7 GHz at +275mV, cache at the time at 4.6 GHz. As soon as selected the built in blend test, boom, error. I kept backing off until I was at stock (4 GHz), with CPU voltage offset still in place, and still got errors instantly. Removing the voltage offset, it took over 6 minutes before I got an error. I'm not sure what to make of it at the moment. The instant fails were at 400k FFT, and the 6 minute fail was at 480k. These are interesting as that's just under 4MB per thread, so it would be a mix of using data from cache and ram, as opposed to bigger FFT sizes that would be more leaning to the ram, or smaller ones more leaning to the cache. Is the extra voltage to the CPU causing it to fail faster? Why? I guess I'll work back now, and see if the earlier apparently stable conditions would be under this p95 loading.

 

[mackerel]

Done some more with voltages. In short, increasing SA to 1.20v helped a lot, but wasn't 100% stable. Going on to 1.24v didn't help further. I also tried vmem to 1.40 and IO to 1.14 neither of which helped. It seems throwing more voltage at it at this point isn't helping. Question is, where next? Relax timings? Which ones? Given SA helped, does that point more towards IMC than ram itself? What can help there?

 

[Woomack]

I'm not sure if trying to make it work at 3733+ is worth it. At least for what you are using it, it can be waste of time.
For some reason all memory manufacturers are selling 3733+ kits at more relaxed timings. It depends on many factors if memory is stable. I wasn't testing CL so I don't know how it's acting at higher voltages but usually too high and too low voltage is only making things worse. can try SA and IO up to 1.3V but I don't know if it's going to help.
I don't know about any issues with tRAS but as I said, I had no chance to play with CL and if I won't get any mobo for review then I won't even touch it.
I would try high tRFC like 600+ and CL=wCL ( twCL, usually at the end of main timings list, writes CL ).

 

[mackerel]

I got it stable in the end. Had to relax tRAS slightly. The ram current setting is 3866 17-17-17-38 1.35v, IO 1.10v, SA 1.20v. Everything else on auto, which for tRFC is 528, and I had it running at 400 previously but haven't gone back to validate with current other settings. Stability here is defined as nearly 3h of Prime95 custom 400k-4096k, and 1h40+ Aida64 mem+cache. I think I'll make P95 400k FFT my new "quick check" test as it provoked instability where aida64 seemed fine.

"for what you are using it"... I'm curious what you think that is I'm doing it mainly for self education in the dark arts of ram OC. As a side effect it might give me a fraction advantage in hwbot submissions, although I think I'm going to be more held back by core for now. Also my eventual 24/7 usage will be in tasks which I know are highly sensitive to ram bandwidth. With the above ram settings, I've also tried a little CPU overclock, and had it running at 4.7 GHz core, 4.4 GHz cache, Prime95 blend ok for 50 minutes while I was doing other things. Actually cache is a little disappointing. It isn't stable with P95 at 4.5 GHz+, yet for other benching I've had it running 4.6 without problem. It is near instant BSOD with P95, not P95 rounding error or other crash.

I'm still debating, if I should try to dial in further optimisations at 3866, or I could try my luck and have another go at getting 4000 stable now that I'm picking up some tricks.

 

[mackerel]

I'm still tinkering with the ram, and not finished yet. Here's an intermediate result - haven't compared it to others, will hunt out Woomack's other posts later.

 

[UltraTaco]

Mama mia! 50,000mb/s ram! That's insane! Mine is like 15,000. I guess I could try tightening some timings to reach 50 mark..

Interesting read.

 

[||Console||]

Sure beats up my ddr3 =)

 

[mackerel]

Hit my first setback tonight... was working on micro-managing more timings. Just as I was about to do a long stability test I got an error in Prime95. Doh! Back out the change, error still there. Back out that... suffice to say I was almost back to how I was at the start of the session before I didn't get the error again. At least, not quickly. I thought things were going rather well. Can't remember name of the timing but I dropped it a lot and hit the mobo allowed minimum. The interesting thing was, when I was backing out changes, I found it either happened right at the start of a P95 stress, or it would go on for minutes ok. I wonder if that points to a wobbly patch of ram that gets randomly hit on each run? I was doing spot checks as I made the changes, and by luck maybe I missed the instability, hence continued on.

So right now, I need to establish a stable state to work from. I'll leave P95 running for a bit, before doing likewise with Aida64. Once I'm happy, at least I can re-apply the apparent ok settings I found afterwards and just not so aggressively tweak the current one.

Note I found overclocking the CPU and cache helped the ram benchmarks. For the tweaking testing I'm running fixed 4 GHz for both, and I'm getting in the ball park of 53 GB/s read, 58 GB/s write, and latency around 40ns. Compared to last night's intermediate result, looks like write has gone up although I wont know about reads until I re-OC.

 

[||Console||]

What you are going through is why I don't tweak ram timings anymore . Not worth the time invested for what I do(I will still OC men but just speed) for 24/7.

On a side note my final stress test is always csgo it crashes on things that pass p95/aida64/real bench. So if it's not stable enough for a few rounds not worth hours testing.