I went DEEEEP on RAM overclocking today... advice & discuss?

[HoneyBadger84]

So this is all Buildzoid's fault... in a good way, so far.  He posted this video: 

and since it was done with my... well, my exact setup minus the RAM, I decided to give it a try myself.  Now before you go "WHY THE..." I'm not pushing anywhere NEAR as hard as he is or as high. I am just very picky about numbers and I very much wanted to see if my RAM could run 4000MHz at decent timings without murdering it with voltage.

 

So, I pretty much started where he started the guide, 1.45V, VCCIO & VCCSA at 1.2V, I let the timings auto, stuck the RAM at 4000MHz, booted up, saw it was at crazy-loose timings (19-22-22 or some crap), rebooted & lowered them, and kept throwing some benchmarks at it while refining things. 

 

Right now here's where I'm at, and I have the CPU at STOCK not 5GHz since I wanted to be sure the RAM is the only thing that could possibly be unstable if something fails/bluescreens, and I also have all 4 sticks installed (since he said for this board/setup, it actually is better to have 4 sticks vs 2):

 

 

And that is at 1.4V on the RAM with the IO/SA still at 1.2V, the majority of those timings are still auto.  Also some things in the screenshot aren't quite showing right, my Northbridge/Uncore is at 4.3GHz (stock).  Also for anyone seeing rising RAM temps, if you have RGB units, disable the RGB lighting while you're testing things, as that will lead to cooler temperatures (not a huge difference, 1-2C, but still, every little bit helps according to Buildzoid).

 

Couple of questions for anyone that's also been there done that or is doing this, what do you use to test your RAM?  Just MemTest, or do you also use Prime95 etc?  

 

What timings are you guys running at what clocks, do you have any obvious recommendations where I could improve further? I'm thinking about pushing the tRFC down under 400, but I don't know how low I should be going at this speed.  I did try 16-16-16-36 & it booted but I could tell the system wasn't very happy with it right away by the weird results from CineBench & what little else I ran before reverting to 18-18-18.

 

Oh, also, when I say I went deep: I've been at this for like 7 hours lol Hence why I'm not seeking suggestions on what else I can try other than "just lowering the timings by one at a time until it fails"

[Bananasplit_00]

try 1T comand rate, those timing seem pretty loose though but i guess at 4x8GB you might not be getting much lower. memtest86 is what i use to check my RAM OCs mainly and once im happy with what i have i try windows and benchmarks and stuff. RAM OCs take forever to do rigth if you want them to be rock solid. My 4x4GB kit of DDR3 is just on XMP atm at 1.45V CL11 2133MHz T1 IIRC

[HoneyBadger84]

1 hour ago, Bananasplit_00 said:

try 1T comand rate, those timing seem pretty loose though but i guess at 4x8GB you might not be getting much lower. memtest86 is what i use to check my RAM OCs mainly and once im happy with what i have i try windows and benchmarks and stuff. RAM OCs take forever to do rigth if you want them to be rock solid. My 4x4GB kit of DDR3 is just on XMP atm at 1.45V CL11 2133MHz T1 IIRC

I don't think I can go to 1T, but I might as well give it a try, I'll reboot, save the profile to my USB stick & give'er a go.

[porina]

2 hours ago, HoneyBadger84 said:

Couple of questions for anyone that's also been there done that or is doing this, what do you use to test your RAM?  Just MemTest, or do you also use Prime95 etc?  

For a quick confidence check, Aida64 stress test with ram selected for several minutes. Prime95 29.4 large FFT run it for a minute at a time, stop and restart a few times. For whatever reason, I find that either errors quickly, or not at all. I'd do memtest last and infrequently as it is the most time consuming. I'm not that experienced, but proving each change was stable or not was the biggest problem. Each test above does different things, so can find problems others don't. I think P95 large FFT stresses the IMC more than the ram, but it all needs to work together.

 

2 hours ago, HoneyBadger84 said:

What timings are you guys running at what clocks, do you have any obvious recommendations where I could improve further? I'm thinking about pushing the tRFC down under 400, but I don't know how low I should be going at this speed.  I did try 16-16-16-36 & it booted but I could tell the system wasn't very happy with it right away by the weird results from CineBench & what little else I ran before reverting to 18-18-18.

Depends on the ram and mobo... my best was with B-die modules, I think I got to 3866 17-17-17-36 1.35v, IO 1.1, SA 1.20. Dunno if the limit was the modules, the mobo, or my ability. But throwing more voltage at it didn't help.

 

I did my optimisations in two phases, first mostly auto, 2nd I reduced each sub timing, one at a time, to find stability. It gave nice benchmark improvements but not a lot in the real world. And I later found out it wasn't stable after all. I run Prime95 like workloads and if there is a single memory error, they will find it.

[HoneyBadger84]

41 minutes ago, porina said:

For a quick confidence check, Aida64 stress test with ram selected for several minutes. Prime95 29.4 large FFT run it for a minute at a time, stop and restart a few times. For whatever reason, I find that either errors quickly, or not at all. I'd do memtest last and infrequently as it is the most time consuming. I'm not that experienced, but proving each change was stable or not was the biggest problem. Each test above does different things, so can find problems others don't. I think P95 large FFT stresses the IMC more than the ram, but it all needs to work together.

 

Depends on the ram and mobo... my best was with B-die modules, I think I got to 3866 17-17-17-36 1.35v, IO 1.1, SA 1.20. Dunno if the limit was the modules, the mobo, or my ability. But throwing more voltage at it didn't help.

 

I did my optimisations in two phases, first mostly auto, 2nd I reduced each sub timing, one at a time, to find stability. It gave nice benchmark improvements but not a lot in the real world. And I later found out it wasn't stable after all. I run Prime95 like workloads and if there is a single memory error, they will find it.

Right now I'm trying out (still 4000MHz) 17-17-17-36 400tRFC with 1.4V and still 1.2V on SA & IO.  I'll give it a bit of those benchmarks you suggested, it already passed a quick Time Spy, FireStrike & GeekBench test, along with AIDA's Cache & RAM benchmark, but those don't do a whole lot.  I did run Prime95 with Large FFTs for a few minutes as well, mostly to check how the temperature on the RAM was behaving though.

 

Edit: about how long do you let AIDA64's RAM test run?

[porina]

3 hours ago, HoneyBadger84 said:

Right now I'm trying out (still 4000MHz) 17-17-17-36 400tRFC with 1.4V and still 1.2V on SA & IO.  I'll give it a bit of those benchmarks you suggested, it already passed a quick Time Spy, FireStrike & GeekBench test, along with AIDA's Cache & RAM benchmark, but those don't do a whole lot.  I did run Prime95 with Large FFTs for a few minutes as well, mostly to check how the temperature on the RAM was behaving though.

It is very possible to be bench stable, but not 100% stable. That's what caught me out on my tweak everything attempt. Errors were rare enough it was hard to troubleshoot at that point, but it certainly wasn't 100% stable. 

 

3 hours ago, HoneyBadger84 said:

Edit: about how long do you let AIDA64's RAM test run?

It's an eternal question, since we can't ever prove stability. It is only stable until you prove it is unstable. If I was still tweaking, I think it was minutes just to spot any obvious instability. If I wanted to do a deeper verify, an hour or more. I set it and go do something else, keeping an eye on the screen.

[HoneyBadger84]

12 hours ago, Bananasplit_00 said:

try 1T comand rate, those timing seem pretty loose though but i guess at 4x8GB you might not be getting much lower. memtest86 is what i use to check my RAM OCs mainly and once im happy with what i have i try windows and benchmarks and stuff. RAM OCs take forever to do rigth if you want them to be rock solid. My 4x4GB kit of DDR3 is just on XMP atm at 1.45V CL11 2133MHz T1 IIRC

Yeah 1T was no Bueno. Wouldn't even train/post. Worth a try though. 

 

6 hours ago, porina said:

It is very possible to be bench stable, but not 100% stable. That's what caught me out on my tweak everything attempt. Errors were rare enough it was hard to troubleshoot at that point, but it certainly wasn't 100% stable. 

 

It's an eternal question, since we can't ever prove stability. It is only stable until you prove it is unstable. If I was still tweaking, I think it was minutes just to spot any obvious instability. If I wanted to do a deeper verify, an hour or more. I set it and go do something else, keeping an eye on the screen.

I'm hoping that CL17 is "stable", I'm getting inconsistent results though so I think it's not. 17-17-17-36-2T with tRFC at 400, 1.4V at 1.2V SA & IO. I think maybe a slight bump on those SA and IO might help since I'm pushing the memory controller a bit now a wager. Not at home at the moment but I'll test that. The difference between 17s and 18s isn't really that notable. 

 

Last night during a few minutes of AIDA64 ram test the sticks were getting up in to the 35C range according to hwinfo64, so I moved my top intake fan for the VRM up top over a bit so it's blowing in the sticks. That didn't help much so I replaced the 140mm temporarily with a 180mm fan from my old case, seems to work much better, same test for the same length of time only hit 32-33C and cooled off quicker. 

 

There's so little space between the sticks in surprised higher airflow helps at all. 

[HoneyBadger84]

On 2/19/2019 at 5:54 AM, porina said:

For a quick confidence check, Aida64 stress test with ram selected for several minutes. Prime95 29.4 large FFT run it for a minute at a time, stop and restart a few times. For whatever reason, I find that either errors quickly, or not at all. I'd do memtest last and infrequently as it is the most time consuming. I'm not that experienced, but proving each change was stable or not was the biggest problem. Each test above does different things, so can find problems others don't. I think P95 large FFT stresses the IMC more than the ram, but it all needs to work together. 

I'm running an hour long memtest now to give them a thorough torture, is that long enough or should I be letting that just eat it for quite a while to make sure the RAM is stable under even the hottest and most vigorous conditions? I'm running it with 25600MB selected, which puts me at 28GB of the 32GB loaded. Left some head room in case windows gets stupid. 

 

Testing 17-17-17-38-2T 400tRFC 1.4V with 1.2 IO & SA, 14 minutes in with 0 errors. Sticks are quite warm on hwinfo64, 34-36C so far. 

[HoneyBadger84]

Seems it passed indeed.  Screenshot in case I somehow manage to loose my bios profile backup (unlikely but who knows), I still have quite a few timings I know I can probably fiddling with & widdle down, but I'm pretty happy with a tRFC that low as well as fairly tight timings considering my voltage is at a pretty normal 1.4V on the RAM, and both my IO & SA are still in the "normal" range.

 

Also has my "maximum" RAM load temps, 36.5-38.5-38.8-37.3C respectively.  They leveled off at around those temps about 30-35 minutes in.  VRM temp was also laughably-low, but I have a 180mm Silverstone Air Penetrator fan as side-flow for now right over the VRM & RAM.  I think I'm going to put a 200mm Thermaltake fan there eventually, haven't decided yet.

 

[Stormseeker9]

When running memtest86 I do 4 full runs. Takes about 4h I believe in total.

[Domrockt]

are those b-Dies if so they scale well with higher Ramvoltage 

 

higher VRam = lower timings same frequency

 

i consider 1.5V as safe and above 1.6V as scetchy

 

here are some examples (not my systems)

[porina]

For memtest I see that as checking all the ram is working ok, as opposed to the other tests which lean towards general stability than coverage. So for memtest I'd do at a minimum one pass, which quickly covers the whole ram space, and ideally let it run I think it defaults to 4 passes. The later passes are more in depth so can take longer than the 1st pass. Because of this, it is something you can either do overnight, or if you're going away to do something elsewhere. I'd only do the 4 passes as a final check for overall stability for 24/7 use. I wouldn't care if I only wanted benchable settings that wont remain set.

[HoneyBadger84]

13 hours ago, Domrockt said:

are those b-Dies if so they scale well with higher Ramvoltage 

 

higher VRam = lower timings same frequency

 

i consider 1.5V as safe and above 1.6V as scetchy

 

here are some examples (not my systems)

 

 

I don't know if they're B-Die or not, is there some way to check without taking the heat spreader off? Given they didn't lose stability when my dimm readouts in hwinfo64 were showing 38C load temp on 2 sticks, I'm unsure, I've heard that b-die doesn't like high temps but that's at ultra tight timings. 

 

They cooled down fairly quickly thanks to my new airflow setup. I imagine more voltage will mean higher temps as with any hardware, so with that in mind I doubt I'll push much further for 24/7 use. CL17 at 4000MHz with 1.4v seems like a pretty good point to me, especially since these sticks are 3600MHz @ 1.35v CL16 stock. I will probably push further after I get home from work, but I'm not expecting it to work. 

 

I have another PC to assemble so I'll be running further ram testing while I'm doing that. 

[TempestCatto]

15 minutes ago, HoneyBadger84 said:

is there some way to check without taking the heat spreader off?

https://benzhaomin.github.io/bdiefinder/

[HoneyBadger84]

5 hours ago, TempestCatto said:

https://benzhaomin.github.io/bdiefinder/

It would appear so, https://www.newegg.com/Product/Product.aspx?Item=N82E16820232492 that's my current RAM kit and it's listed on that site (model # F4-3600C16Q-32GTZR). Neat. Now I know it can go further if I'm willing to do the fine tuning. 

 

14 hours ago, Domrockt said:

i consider 1.5V as safe and above 1.6V as scetchy

here are some examples (not my systems)

At what point do I consider or need to probably bump my VCCIO & VCCSA voltages do you think? I'm going to push for CL16-16-16-36 tonight and see if I can push my tRFC down as well, since it seems pretty common to run those timings at around 1.4V on the RAM or slightly more. I'm still at 1.2V on both those. Buildzoid's guide kind or glazed over the "when to go up" on that aspect and mostly reinforced what to not go over for 24/7 usage on IO and SA voltage. 

[HoneyBadger84]

Update: 16-16-16-36-2T 350tRFC was no good. Not sure if it was the tRFC change (I doubt it because I've found several people that can run their tRFC that low with voltage in the same range as mine at even higher speeds), or the IO/SA voltages needing a bump, or the RAM itself needing more voltage, but it blue screened when I was doing just regular tasks not even stressing it yet.

 

I'm pretty happy with where I'm at but CL16 would've been spiffy.  I might try 16-17-17-38 tomorrow, right now I'm just gonna relax for a night since it seems that system build I was going to do tonight has been postponed due to parts not arriving.  I can save my next round of tweaking for then since I can stress test while I'm building that & the time will go by quickly.

 

Also, for anyone that's an RGB-head, the GSkill app for RGB lighting on this RAM, while it can conflict with some other RGB software (Gigabyte RGB Fusion is one of them), is absolutely sick in terms of the options you have for RAM lighting control.  You can sync it to music and a whole load of other things as well, for some reason Gigabyte's software today decided it didn't want to see the RAM anymore, which is why I discovered the plus side of GSkill's application.

 

I'm also thinking I may see if I can do something like 3200MHz 12-12-12-28-1T, by the logic of speed/CAS, 3200MHz @ CL12 should actually perform "better" than 4000MHz @ CL17.  I know it's not as simple as that equation, but it's a good starting point.

[HoneyBadger84]

Update: upped the io to 1.25V and SA to 1.22V, ram voltage to 1.45V (I think, will have to check when I'm at home) and 4000MHz CL16-16-16-34-2T tRFC350 has passed all testing so far. Didn't have time to stress test last night, so we shall see. 

 

I also tried CL12 @ 3200MHz and it refused to boot regardless of how I changed other timings it wouldn't train. CL13-12-12 booted but blue screened during a benchmark. But that was at 1.4V. 

 

I'm debating if I want to try lower speed at really tight timings or just keep pushing 4000MHz to see how low I can go on timings without going past 1.45V on the RAM. 

 

[nick name]

I've spoken with several people that run their b-die at 1.5V daily.  Including me.  

[HoneyBadger84]

13 minutes ago, nick name said:

I've spoken with several people that run their b-die at 1.5V daily.  Including me.  

Good to know I have more voltage headroom I'm trying to avoid "the limit" of daily use safety if I can. 

[nick name]

1 hour ago, HoneyBadger84 said:

Good to know I have more voltage headroom I'm trying to avoid "the limit" of daily use safety if I can. 

I'm just glad that you didn't catch on to the fact that I talk to myself.  

[HoneyBadger84]

17 hours ago, nick name said:

I'm just glad that you didn't catch on to the fact that I talk to myself.  

I see what you did there

 

So I tested 4000MHz CL16-16-16-36 tRFC 400 @ 1.45V on the sticks, and 1.25V IO 1.22V SA, and it passed an hour of memtest only to fail Aida64 memory stress test in 5 minutes, it also passed all the benchmarks I threw at it. Suppose I'll have to fiddle with it further. I'm hoping I can get CL16 happy at that speed. 

[nick name]

Have you been checking your RAM temps as you progress?  

[HoneyBadger84]

54 minutes ago, nick name said:

Have you been checking your RAM temps as you progress?  

DIMMs in hwinfo64 topped out during the memtest at 33-35C, so a bit warm but nothing ridiculous. I know that b-die tends to cry a little when it's over 30C at extreme settings but I didn't think that would happen at what I'm trying to run. It was perfectly stable and happy up until that Aida64 test failed. I ran it right after an hour of memtest so the sticks were still "warm". 

[nick name]

46 minutes ago, HoneyBadger84 said:

DIMMs in hwinfo64 topped out during the memtest at 33-35C, so a bit warm but nothing ridiculous. I know that b-die tends to cry a little when it's over 30C at extreme settings but I didn't think that would happen at what I'm trying to run. It was perfectly stable and happy up until that Aida64 test failed. I ran it right after an hour of memtest so the sticks were still "warm". 

Actually, mine didn't lose stability at the edge until 40*C+ in the summer.  

 

And that Aida64 -- does that pass without the memory overclock?  Perhaps it's a combination of the CPU and RAM.  

[HoneyBadger84]

46 minutes ago, nick name said:

Actually, mine didn't lose stability at the edge until 40*C+ in the summer.  

 

And that Aida64 -- does that pass without the memory overclock?  Perhaps it's a combination of the CPU and RAM.  

I'm running all memory OCs with the CPU at stock to eliminate it as a potential factor (unless it's lack of IO/SA voltage), and everything passes with the RAM at bone stock, made sure of that before I started goin bonkers with the overclocking.

 

It passes an hour of Aida64 with just memory checked as well as all 4 cpu/fpu/cache/memory checked at 4000MHz CL17 @1.4V, so it's either a secondary timing, not enough voltage, or something I'm missing. Just very odd that CL16 @ 1.45v passed memtest but failed elsewhere. But, that's why we do more than one check I suppose. 

 

I'll probably try a slight voltage bump again. I imagine 4000MHz @ CL16 isn't exactly easy to get settled down. Wish there was a bit more guidance on when it's prudent to raise VCCSA and VCCIO and by how much.  Could be any of the 3 voltages need a bit more, as it's probably pretty close to stable.

 

Are you using 1.5V on the sticks for the RAM settings on your signature?