Comprehensive Memory Overclocking Guide

[mrthuvi]

My gigabyte board has a setting called Enhanced Performance for memory which reduce latency a little bit but reduce read/write quite a bit.

 

Stock:

 

Enhance Performance: 

 

 

Which one is better? @MageTank

[MageTank]

38 minutes ago, mrthuvi said:

My gigabyte board has a setting called Enhanced Performance for memory which reduce latency a little bit but reduce read/write quite a bit.

 

Stock:

 

Enhance Performance: 

 

 

Which one is better? @MageTank

All it did was change CAS latency from 16 to 15, lol. It also changed to single channel, as you can see in your screenshot. Avoid that "enhanced" mode like the plague. 

[mrthuvi]

Thanks, this board is expensive but quite shitty to overclock, frequently I enter a value but boot to different value. 

[XenosTech]

Check your dms @MageTank senpai q.q

[Carclis]

@MageTank

So I've been playing around with a new kit of memory and I've managed to get it stable with 4*8GB dimms at 19-19-19-38 4000Mhz through Prime95 testing for 8+ hours multiple times. I'm running into an issue where it's perfectly stable for use but will fail to boot resulting in a 55 code about 25% of the time. Any ideas on what this might be or if you've experienced something similar?

[MageTank]

8 hours ago, Carclis said:

@MageTank

So I've been playing around with a new kit of memory and I've managed to get it stable with 4*8GB dimms at 19-19-19-38 4000Mhz through Prime95 testing for 8+ hours multiple times. I'm running into an issue where it's perfectly stable for use but will fail to boot resulting in a 55 code about 25% of the time. Any ideas on what this might be or if you've experienced something similar?

I have experienced something similar on my ASRock board, known as a "slow post". The fix was to manually dial in every timing (with MRC Fast Boot completely disabled) and never had the issue again. I think it's caused by your board trying to retrain values at every post, causing it to hang when the values don't quite work. I highly recommend using RTL Init to prevent your RTL/IO-L values from training poorly. RTL Init's formula is: IO-L + IO-L Offset + CAS Latency (x2) + 10. So if your IO-L is 7, the offset is 21, and CAS Latency is 19, your RTL Init would be: 7+21+38+10=76. 

 

You will need to find your RTL/IO-L values to be certain of the exact number, but that's how you do the math to get the proper value. This will prevent your RTL/IO-L from training outside of it's current values, which is great from a stability standpoint. 

[Carclis]

15 hours ago, MageTank said:

I have experienced something similar on my ASRock board, known as a "slow post". The fix was to manually dial in every timing (with MRC Fast Boot completely disabled) and never had the issue again. I think it's caused by your board trying to retrain values at every post, causing it to hang when the values don't quite work. I highly recommend using RTL Init to prevent your RTL/IO-L values from training poorly. RTL Init's formula is: IO-L + IO-L Offset + CAS Latency (x2) + 10. So if your IO-L is 7, the offset is 21, and CAS Latency is 19, your RTL Init would be: 7+21+38+10=76. 

 

You will need to find your RTL/IO-L values to be certain of the exact number, but that's how you do the math to get the proper value. This will prevent your RTL/IO-L from training outside of it's current values, which is great from a stability standpoint. 

This seems to do the trick. Thanks for that.

[MageTank]

So, I just recently got hired for a job that will help me obtain more hardware at decent prices. Once I secure a couple paychecks, I'll invest in either a Ryzen or Threadripper build, so that I can personally update the Ryzen section of this guide with more information. Right now, the only Ryzen experience I have is through my friends Ryzen 1600 that I visit every now and then to OC on. 

 

I want to invest in budget-oriented hardware, just so that I can show people that you CAN overclock on cheaper hardware, so long as you know what you are doing, and know what compromises to make in order to get the best performance possible. I will also probably compare the difference between a budget board and enthusiast level board when it comes to memory overclocking, and the difference it makes for framerates when comparing FPS. Stay tuned for that information in the future. 

[Jade]

Oh, right - I asked for help here a month or two back, and since I keep pinging Mage about RAM, I might as well post results. I switched out the motherboard (kept the same dimms) & RMA'd the old CPU (had segfaulting problem) now they're plenty happy at 3200MHz with somewhat decent timings.

[MageTank]

1 minute ago, Jade said:

Oh, right - I asked for help here a month or two back, and since I keep pinging Mage about RAM, I might as well post results. I switched out the motherboard (kept the same dimms) & RMA'd the old CPU (had segfaulting problem) now they're plenty happy at 3200MHz with somewhat decent timings.

I must say, I am quite impressed with your read and write bandwidth. Peak theoretical bandwidth of 3200mhz is 51.2GB/s. You are achieving almost 99% read efficiency (nearly unheard of), and 98% write efficiency. Copy efficiency is difficult to achieve, as it basically depends on how many ranks you have (so technically, 2DPC multi-rank kits would boost copy at the cost of clocks), and take it from a man that sincerely tried, you cannot tame the latency of the Ryzen memory controller. All in all, this is one of the tightest 3200mhz OC's I have ever seen on Ryzen. Thoroughly impressed. 

[WHWidjaja]

Just wanted to share my result. G.SKILL TridentZ DDR4 3200 (PC4 25600) F4-3200C14D-16GTZKW @ DDR4 3466 CL14-14-14-48 1T

 

[porina]

On 01/05/2017 at 3:41 PM, MageTank said:

As for why we use the settings above, allow me to explain. 512k-1024k is hard on the IMC and IO lanes. 2048k+ is hard on your ram. By setting the range at 512-4096, we not only stress the IMC and IO Lanes, we also stress the memory itself. Be warned: 1344k and 2688k are also included in this range, and are the hardest stress on vCore. If your CPU is unstable by any means, it will fail this, and will likely hold you back on memory overclocking. Always make sure your CPU is 100% stable before attempting memory overclocking. The less variables involved, the better. For those of you with Haswell, and worried about that old myth of Prime95 killing CPU's, understand this. This range lacks 448k, which was the hardest FFT to test on FIVR. You should be fine here.

I've been doing my first really serious ram OC today, and I've had some interesting observations when using Prime95 and various FFT sizes.

 

Some history: I'm using an i3-8350k (stock), Asrock Z370 Pro4, TridentZ 3000C14 (B-die) 2x8GB. My previous "stable" condition was 3600C16 (1.35v, IO 1.05v, SA 1.15v), and today I finally managed to go beyond that. For reasons unknown, it just wont boot at 3700 ram regardless what I do. It worked 1st time at 3733, and can only speculate it hates 100 MHz strap for some reason. I'm currently what you might call pseudo-stable at 3866, and have also booted at 4000 but haven't tried to stabilise that, concentrating on 3866 for now. My test setting is 3866 17-17-17-36 with everything else on auto. To stabilise it, I increased IO to 1.10v. The other two didn't make a difference. My chosen stability test was aida64 stress test with cache and ram selected only. In the past I found memtest to be relatively slow and didn't really pick up errors that would be quickly found with aida64. With CPU at stock, this configuration passed over an hour.

 

I rediscovered this guide, and played a little on tRFC (400 set worked, 300 no boot) and tREFI (15000), although I didn't test those settings very deeply. At the time, I tried the 512k-4096k and that was ok for 30 minutes or so.

 

I had observed that my CPU cache was running at "only" 3.7 GHz stock. I was thinking, was this holding back the ram which was actually at higher speed? Past overclocking had shown 4.6 GHz+ was ok, so I just set it to 4.0 for now without touching voltage. That gave a bit of a boost to aida64 bandwidths. I thought I might as well OC the CPU as well, which was bench stable at +275mV for 5.0 GHz for non-AVX, but I never found the AVX offset. At 128k FFT (so it wouldn't need ram and only stress core), it was fine to 4.7 GHz. I didn't test higher, as I think previous attempt at 4.8 failed.

 

To recap, 128k FFT seemed fine at 4.7 GHz, 512k-4096k seemed ok at stock. As I was going to do something else, I thought I'd just put p95 blend on to leave running... it instantly errored. Ok, back off the overclock. Still errors. Back to stock clock (but still with extra voltage) - instant error. Remove extra voltage, now it took around 6 minutes before erroring. Extra CPU voltage seemed to make it error faster? As an observation, the instant fails were at 400k FFT, and the 6 minute fail was at 480k FFT, so they're all in a similar ball park, sizes which are bigger than the CPU cache so it will have some use of ram. Is this pointing to an IMC instability more so than the ram? I could play more with IO/SA voltages, but haven't done so yet.

 

In case it makes any difference, I'm using Prime95 29.3. It might choose to use slightly different optimisations than other older recent versions as there was work done to use more optimal FFTs for performance.

 

[MageTank]

5 hours ago, porina said:

I've been doing my first really serious ram OC today, and I've had some interesting observations when using Prime95 and various FFT sizes.

 

Some history: I'm using an i3-8350k (stock), Asrock Z370 Pro4, TridentZ 3000C14 (B-die) 2x8GB. My previous "stable" condition was 3600C16 (1.35v, IO 1.05v, SA 1.15v), and today I finally managed to go beyond that. For reasons unknown, it just wont boot at 3700 ram regardless what I do. It worked 1st time at 3733, and can only speculate it hates 100 MHz strap for some reason. I'm currently what you might call pseudo-stable at 3866, and have also booted at 4000 but haven't tried to stabilise that, concentrating on 3866 for now. My test setting is 3866 17-17-17-36 with everything else on auto. To stabilise it, I increased IO to 1.10v. The other two didn't make a difference. My chosen stability test was aida64 stress test with cache and ram selected only. In the past I found memtest to be relatively slow and didn't really pick up errors that would be quickly found with aida64. With CPU at stock, this configuration passed over an hour.

 

I rediscovered this guide, and played a little on tRFC (400 set worked, 300 no boot) and tREFI (15000), although I didn't test those settings very deeply. At the time, I tried the 512k-4096k and that was ok for 30 minutes or so.

 

I had observed that my CPU cache was running at "only" 3.7 GHz stock. I was thinking, was this holding back the ram which was actually at higher speed? Past overclocking had shown 4.6 GHz+ was ok, so I just set it to 4.0 for now without touching voltage. That gave a bit of a boost to aida64 bandwidths. I thought I might as well OC the CPU as well, which was bench stable at +275mV for 5.0 GHz for non-AVX, but I never found the AVX offset. At 128k FFT (so it wouldn't need ram and only stress core), it was fine to 4.7 GHz. I didn't test higher, as I think previous attempt at 4.8 failed.

 

To recap, 128k FFT seemed fine at 4.7 GHz, 512k-4096k seemed ok at stock. As I was going to do something else, I thought I'd just put p95 blend on to leave running... it instantly errored. Ok, back off the overclock. Still errors. Back to stock clock (but still with extra voltage) - instant error. Remove extra voltage, now it took around 6 minutes before erroring. Extra CPU voltage seemed to make it error faster? As an observation, the instant fails were at 400k FFT, and the 6 minute fail was at 480k FFT, so they're all in a similar ball park, sizes which are bigger than the CPU cache so it will have some use of ram. Is this pointing to an IMC instability more so than the ram? I could play more with IO/SA voltages, but haven't done so yet.

 

In case it makes any difference, I'm using Prime95 29.3. It might choose to use slightly different optimisations than other older recent versions as there was work done to use more optimal FFTs for performance.

 

Do you happen to know the error? Is it rounding up or down? Once I get this information, I'll be able to better assist you with making it stable. Prime95 is pretty tricky to diagnose without knowing exactly what is rounding, and the direction it's rounding to. 

[porina]

16 hours ago, MageTank said:

Do you happen to know the error? Is it rounding up or down? Once I get this information, I'll be able to better assist you with making it stable. Prime95 is pretty tricky to diagnose without knowing exactly what is rounding, and the direction it's rounding to. 

Here's some of them:

 

Quote

FATAL ERROR: Rounding was 0.5, expected less than 0.4
Hardware failure detected, consult stress.txt file.
FATAL ERROR: Rounding was 8.651888561e+018, expected less than 0.4
Hardware failure detected, consult stress.txt file.
FATAL ERROR: Rounding was 2.818204458e+019, expected less than 0.4
Hardware failure detected, consult stress.txt file.
FATAL ERROR: Rounding was 3.686959539e+018, expected less than 0.4
Hardware failure detected, consult stress.txt file.
FATAL ERROR: Rounding was 4.01357945e+018, expected less than 0.4
Hardware failure detected, consult stress.txt file.

Unfortunately the log doesn't include the FFT size that failed, but most of these would be 400k.

 

Since it was unstable faster, I tested the 3866 configuration, without rFRC + tREFI changes (now auto), with CPU+cache at 4000 and +275mV offset, with memtest86. It passed 3 passes without error. This same configuration instantly gives error in Prime95.

 

To me fast reproduceable errors are good in a way, at least it is easy to make a change and see quickly if it has made any difference. I'm going to try playing with the voltages a bit more now...

 

Edit: bumping up IO to 1.14v didn't help, returned that to 1.10v and increased SA to 1.20v. That didn't instantly give an error, so will leave that running for now.

Edit 2: run with SA at 1.20v eventually gave a rounding error 0.4999 at 1280k FFT. Further increase SA to 1.24v and ram to 1.40v, error 0.4999 at 672k FFT.

Edit 3: increase IO again to 1.14v, error 0.4999 at 1920k. Running out of voltages to tinker with.

Edit 4: I think I cracked it. Relaxing tRAS from 36 to 38, has passed about couple hours of Prime95 so far. Current test settings are 3866 17-17-17-38 1.35v, IO 1.10v, SA 1.20v. This ram tinkering stuff... so many variables!

[porina]

 

After best part of a week tinkering on and off, above is where I'm at. How's it looking? At this point I have fiddled with every primary/secondary/tertiary timing I can get my hand on, although I can't say every single one has been pushed as far as they can yet. There's at least a couple I could be more aggressive on.

 

I have a feeling this is something I'll only ever do once. The time cost is rather insane, and once is enough. On the weekend I think I'll do a comparison of a few states of ram, like 2133 SPD, 3000C14 XMP, 3600C16+auto, 3866C17+auto, 3866C17+manual. See how things change at each step, both synthetics and some TBD sample workloads.

[MageTank]

31 minutes ago, porina said:

 

After best part of a week tinkering on and off, above is where I'm at. How's it looking? At this point I have fiddled with every primary/secondary/tertiary timing I can get my hand on, although I can't say every single one has been pushed as far as they can yet. There's at least a couple I could be more aggressive on.

 

I have a feeling this is something I'll only ever do once. The time cost is rather insane, and once is enough. On the weekend I think I'll do a comparison of a few states of ram, like 2133 SPD, 3000C14 XMP, 3600C16+auto, 3866C17+auto, 3866C17+manual. See how things change at each step, both synthetics and some TBD sample workloads.

Sorry for the inactivity on my part, this Cyber Monday (week) sales has kept me extremely busy at work lately. Looking at the results above, you are hitting 88% read efficiency, 92% write efficiency, and roughly 83% copy efficiency. This is by no means a bad thing, especially once you consider your latency is 38ns. Anything below 40 is considered fantastic for DDR4. Another trick, if you can get your cache to run stable at higher clocks, you will be able to shave off even more latency as well, so that is something to consider in the long run if you don't mind the additional stress testing involved with stability validation.

 

If you are able to tighten tRFC any, and raise tREFI up as well, you can pull off even lower latency, it just makes it more difficult to maintain stability over a long period of time. I'd say for a 24/7 overclock, this destroys almost any XMP profile one could ever buy in terms of latency:bandwidth. It certainly gets my seal of approval, if that means anything at all, lol. 

[porina]

 

8 hours ago, MageTank said:

Another trick, if you can get your cache to run stable at higher clocks, you will be able to shave off even more latency as well, so that is something to consider in the long run if you don't mind the additional stress testing involved with stability validation.

44x is the fastest I'm p95 stable at. I can run 46x for non-AVX workloads, but there's no AVX offset for cache (yet).

 

8 hours ago, MageTank said:

If you are able to tighten tRFC any, and raise tREFI up as well, you can pull off even lower latency, it just makes it more difficult to maintain stability over a long period of time.

Those are two of the settings I haven't really pushed hard yet. tRFC was bought down from I think 526 to 380 currently. I know 300 doesn't boot, so there is still a gap untested. rREFI was increased from auto 11000-ish to 15000 currently as a "safe" setting based on your guide. I've only briefly tried 16000 and again could push further.

Quote

I'd say for a 24/7 overclock, this destroys almost any XMP profile one could ever buy in terms of latency:bandwidth.

Haven't verified this in depth yet. A few settings less had 4 passes of memtest86 as well as some time in p95 and aida64. Shall try to give the above two a bit more tweaking and I'll post a full list of settings if believed stable.

Quote

It certainly gets my seal of approval, if that means anything at all, lol. 

I'll take it!

[Clyne]

Board somehow doesn't like 3600 unless I adjust tRDWR to match tCL number....I'm sure gonna fry the IMC soon at that voltage.

[MageTank]

20 minutes ago, Bubblewhale said:

Board somehow doesn't like 3600 unless I adjust tRDWR to match tCL number....I'm sure gonna fry the IMC soon at that voltage.

 

That is extremely impressive. I wouldn't worry about 3600, your 3485 is nearly competing with my extremely tight 3600 C14 setup. In fact, if you can push that cache clock any higher, your latency would certainly compete. Also, if you can get tRFC down to 270-320, you'd see a pretty big improvement in latency as well. 

[Clyne]

3 minutes ago, MageTank said:

That is extremely impressive. I wouldn't worry about 3600, your 3485 is nearly competing with my extremely tight 3600 C14 setup. In fact, if you can push that cache clock any higher, your latency would certainly compete. Also, if you can get tRFC down to 270-320, you'd see a pretty big improvement in latency as well. 

I have gotten 4.8-4.9GHZ cache stable, but that was with around 1.44v.  Think I'll go around with a slight increase in vcore for 4.5-4.6 cache.

[Jade]

On 9/27/2017 at 7:54 PM, MageTank said:

I must say, I am quite impressed with your read and write bandwidth. Peak theoretical bandwidth of 3200mhz is 51.2GB/s. You are achieving almost 99% read efficiency (nearly unheard of), and 98% write efficiency. Copy efficiency is difficult to achieve, as it basically depends on how many ranks you have (so technically, 2DPC multi-rank kits would boost copy at the cost of clocks), and take it from a man that sincerely tried, you cannot tame the latency of the Ryzen memory controller. All in all, this is one of the tightest 3200mhz OC's I have ever seen on Ryzen. Thoroughly impressed. 

Switched kits again, so thus, I return! I'm on 2 sets of 2x8GB TridentZ RGB 3200 14-14-14-34 sticks, now. Haven't had too much time to tinker, so this is the best I could do... will probably come back and play with it some more after BIOS 3501 makes it to stable. Good news is that this has been stable for over a week 

[7nationarmy]

Hey guys, I'm new to overclocking in general, and I am wondering if this is a good result for my kit (Hynix M-die 16GB 3000/16-18-18-38). I used Ryzen DRAM Calculator I found at overclock.net (which I am not sure if that is the best way to OC ram) and the results have been much more stable than my first attempt (just changed the speed and voltage, timings all on auto). CL20 seems pretty high, but after trying out 2933, 3200 and 3466, this gives the lowest latency and highest read/write/copy. Appreciate the help and thanks in advance!

 

[simo FLC]

Thanks, great Job, This post clear my mind!

[Gre4ka]

MageTank

I'm very happy that I find your post and some explanations regarding memory overclocking & stability testing.

Thread seems to be dead, but I really want to try

 

I moved to new platform a few weeks ago from 1366 socket with 6-core westmere. I missed all intel cpu's, their overclocking tips&tricks, ddr4 etc, while I was happy with my build for about 7 years. I read a lot regarding overclocking on new platforms, safe voltages, bios tweaking etc. I dellided my CPU myself and successfully from first try. But I really stuck with memory overclocking and I can't understand where the problem is really - not stable CPU (but now it's can be broken on 3 parts - non-avx, avx/2, fma3), not stable memory (OR specifical AVX loads on/from it) or cache.

 

It was very hard for me to stabilize memory on it's vendor specs. 3200 14-14-14-34 1.35v

I made a lot of errors following terrible guides like "put your tRFC to 360, tREFI to 65535 and you're done! +10% in bandwidth and ~performance and you don't need to touch anything else". Yeah, taking this 2 timings to such levels gives a lot of performance, latency decreases, but of course it's not "free" as it becomes unstable. I was really frustrated when got errors in memtest from HCI after a hours of test (for example 200-300% coverage). Then I set all this to auto and played with RTL (again following terrible guides).

 

Finally I stuck at 3200, because my memory kit or my CPU IMC don't allow me to play any higher. Maybe I missing something or so.

Now I'm testing 5GHz overclock with memory together. I'm using prime95 (latest version) with adding options to local.txt for desired test (no avx/fma, avx, but no fma, or fma)

And I still can't understand where the problem comes from.

 

If you have some free time I may post you private message with more detailed explanations and description of settings and results (errors). Or we can continue here or I can make a new thread.

 

Thanks in advance!

 

[MageTank]

On 10/30/2018 at 6:20 AM, Gre4ka said:

MageTank

I'm very happy that I find your post and some explanations regarding memory overclocking & stability testing.

Thread seems to be dead, but I really want to try

 

I moved to new platform a few weeks ago from 1366 socket with 6-core westmere. I missed all intel cpu's, their overclocking tips&tricks, ddr4 etc, while I was happy with my build for about 7 years. I read a lot regarding overclocking on new platforms, safe voltages, bios tweaking etc. I dellided my CPU myself and successfully from first try. But I really stuck with memory overclocking and I can't understand where the problem is really - not stable CPU (but now it's can be broken on 3 parts - non-avx, avx/2, fma3), not stable memory (OR specifical AVX loads on/from it) or cache.

 

It was very hard for me to stabilize memory on it's vendor specs. 3200 14-14-14-34 1.35v

I made a lot of errors following terrible guides like "put your tRFC to 360, tREFI to 65535 and you're done! +10% in bandwidth and ~performance and you don't need to touch anything else". Yeah, taking this 2 timings to such levels gives a lot of performance, latency decreases, but of course it's not "free" as it becomes unstable. I was really frustrated when got errors in memtest from HCI after a hours of test (for example 200-300% coverage). Then I set all this to auto and played with RTL (again following terrible guides).

 

Finally I stuck at 3200, because my memory kit or my CPU IMC don't allow me to play any higher. Maybe I missing something or so.

Now I'm testing 5GHz overclock with memory together. I'm using prime95 (latest version) with adding options to local.txt for desired test (no avx/fma, avx, but no fma, or fma)

And I still can't understand where the problem comes from.

 

If you have some free time I may post you private message with more detailed explanations and description of settings and results (errors). Or we can continue here or I can make a new thread.

 

Thanks in advance!

 

Greetings, glad you were able to find the guide. The thread has definitely been neglected due to my heavy work schedule, but I'll gladly assist you in any way that I can. 

 

In regards to your specific situation, I think it would be best to start with the basics, and try to purge some of the information that may not be in your best interest. Luckily, it sounds like you already have a good foundation with your XMP profile if it's letting you remain stable with RAM while running your CPU overclock. The trick here will be tweaking RAM in a way where you get meaningful performance without compromising on stability. To do so, you will need to make slow, incremental adjustments. Start with setting static voltages for VDIMM, VCCIO and VCCSA. For VDIMM, we can start at 1.39v and work our way down from there once we find a nice solid clock speed. Do not worry about the voltage, it's harmless on even the cheapest Intel boards that I've tested. Set VCCIO and VCCSA to anywhere between 1.1-1.14v. You honestly do not need to go any higher once you tweak your tertiary timings correctly. From these static voltages, dial in static primary timings. In this case, keep 14-14-14-28, and set your command rate to 2. We can always change command rate later as well. From these static primary timings, slowly increase your memory strap until you run into instability. When you do, stop tweaking the clock speed, and start working on timings. Start with 13-14-14-28-CR2. You can also try 13-15-15-30-CR2 as well, which might work out a little better depending on the specific kit.

 

Once you've found a solid clock speed and set your primary timings, you can reference the rest of the guide to learn how to adjust secondary/tertiary timings. Let me know if you have any questions.