Did I manage to speed up my ram or not?

Jump to latest Follow Reply
yd

yd

Ars Legatus Legionis
20,764
Subscriptor++
  • #1
    So I bought 4 16gig sticks of G.skill Trident Z ddr4 3600Mhz cas 17 ram.

    When I initially built the desktop it went into, it wouldn't post if I set the bios timing to 3600mhz. I gave up with 'auto' and let it be.

    So I finally did the bios upgrade. Touched nothing, its all nice and stable but the ram speed was showing 2133Mhz in Windows 10 task manager.

    So now I changed it to 3600Mhz in the bios and in Windows, I am seeing 3600Mhz as per here....

    0EcXdlj.png


    However, if I run CPU-Z I see this for 'memory' tab and 'speed' tab so I am not sure if that is really working at 3600Mhz

    WqHtfrH.png


    I also ran HWinfo and again, not sure if its running at 3600Mhz based on it either

    s7nQ5Fb.png



    So, have I actually accomplished anything?
     
    whoisit

    whoisit

    Ars Tribunus Angusticlavius
    6,565
    Subscriptor
  • #3
    Here's CPUZ and Task Manager on my system running DDR4-3200

    TaskMgr.png

    CPUZ.png

    CPU-Z is reporting the right frequency. Remember DDR is double pumped. It does two reads per clock cycle. So take the marketing number and divide it by 2 to get the actual frequency it's running at. That CL of 26 is adding a lot of latency though. As said above, try enabling XMP and see if you can get the better timings that should be going with the clock rate. If you are running a newer Intel CPU (11th Gen and up. not applicable to AMD), maybe try putting the CPU in Gear 2 if XMP is not stable in Gear 1.
     
    Semi On

    Semi On

    Senator
    89,415
    Subscriptor++
  • #6
    yd said:
    There is no XMP though - its AMD so I don't see a setting that is 'xmp' on a Asus x570-e mobo (it was easier on the wife's desktop and the htpc as they were Intel based and xmp made it automagic).
    Click to expand...
    XMP is Intel. Look for DOCP in the BIOS. I think that's Asus's self branding for Expo (and XMP? I haven't owned an Intel based Asus board in a while).
     
    • Like
    Reactions: continuum
    yd

    yd

    Ars Legatus Legionis
    20,764
    Subscriptor++
  • #7
    Ok, well there was DOCP available along with 'auto' and 'manual' under AI Overclocking (which I hadn't used as I thought that did like the whole system). Well I used it and it looked to do ram similar to what happened on the Intel boards so went with it. No post.

    So I switched back to just auto to get back to where I was pre changing the one line to get memory to 3600Mhz (so auto across all options) and posted as before. Timings better, speed back to 2133Mhz in task manager and details in the other two seem to show better timings - I never touched timings, just speed so have no idea why the latency numbers changed.

    jWmA5j7.png


    YNdcM5j.png
     
    whoisit

    whoisit

    Ars Tribunus Angusticlavius
    6,565
    Subscriptor
  • #10
    Last edited:
    • Like
    Reactions: continuum
    Semi On

    Semi On

    Senator
    89,415
    Subscriptor++
  • #11
    If what @whoisit suggested doesn't work (I think that's a good place to start):

    Try pulling one of the pairs, boot to bios, set DOCP and see if it boots all the way. Put the other sticks back in to the original spot and pull the ones that posted. If it fails, turn it off and put the sticks that booted successfully into that slot and try again. You may have a pair that's not up to the task or the motherboard can't handle that speed on both banks.

    I have a pair of DIMMs that would only work in one pair of slots but not the other. Later, I ordered another pair and it worked in the slots that didn't work prior. No idea if it was a bios improvement or just the second pair of DIMMs were better.
     
    C

    cerberusTI

    Ars Tribunus Angusticlavius
    6,449
    Subscriptor++
  • #12
    I would specifically check that the voltage it indicates under the XMP profile is set when you use DOCP, and if not, set it.

    Also, be aware that four modules does not always work at the rated speed for two. You may need to further increase voltage to do that (which has some risk), or not be able to do so at all.
     
    Last edited:
    • Like
    Reactions: continuum
    yd

    yd

    Ars Legatus Legionis
    20,764
    Subscriptor++
  • #14
    So, think we have success. In the bios screen I enabled under 'AI Overclock Tuner' the DOCP option. That though put the memory to DDR 3603 17 19 19 19 39 and the dram voltage from auto to 1.35. However, as we kinda know, 3600 not likely to work (good call on that chip question whoisit!). I had no idea the chip would limit the ram (so I bought more chip than I needed as I probably shoulda got the 12 core anyways and the 16 core I got ended up with me getting and paying for faster ram than I can use doh).

    So I manually memory frequency option back DOWN to 3200Mhz and didn't touch anything else.

    Booted up no drama. Windows shows 3200Mhz no drama. Cpuz and HWinfo get us the following.

    qyUQsWP.png


    H1EAoov.png


    So far no crashes but I haven't done anything to push it. To be fair, I doubt I will notice one iota of difference but I did pay for it and I had been meaning to get it sorted - was just quite anxious about faffing around in the bios just for something that likely won't make my experience any way 'quicker' in a way I will feel. I mean, I went from 2133 15 15 15 36 51 to 3200 18 19 19 39 75 - do I win? Oddly, the DOCP in bios seemed to be aiming for 17 and 19 and I got 18 and 19. Close enough?
     
    • Like
    Reactions: Semi On, continuum and BigLan
    B

    BigLan

    Ars Tribunus Angusticlavius
    6,907
  • #15
    Anecdotally, I've been able to tell when one of my kids system's has reset itself to 2400 speed from 3200 (they're otherwise identical 3800x/b450 PCs), but that's when using them back-to-back and I know one has a habit of retraining to the slower speed.

    Hope you get a good placebo effect if nothing else.

    Next, if you've got a high refresh rate monitor check it's not stuck at 60hz :)
     
    • Like
    Reactions: continuum
    yd

    yd

    Ars Legatus Legionis
    20,764
    Subscriptor++
  • #17
    BigLan said:
    Anecdotally, I've been able to tell when one of my kids system's has reset itself to 2400 speed from 3200 (they're otherwise identical 3800x/b450 PCs), but that's when using them back-to-back and I know one has a habit of retraining to the slower speed.

    Hope you get a good placebo effect if nothing else.

    Next, if you've got a high refresh rate monitor check it's not stuck at 60hz :)
    Click to expand...
    Nope, I have a dinosaur for which 60hz is the max....but I am working on that
     
    C

    cerberusTI

    Ars Tribunus Angusticlavius
    6,449
    Subscriptor++
  • #18
    yd said:
    So, think we have success. In the bios screen I enabled under 'AI Overclock Tuner' the DOCP option. That though put the memory to DDR 3603 17 19 19 19 39 and the dram voltage from auto to 1.35. However, as we kinda know, 3600 not likely to work (good call on that chip question whoisit!). I had no idea the chip would limit the ram (so I bought more chip than I needed as I probably shoulda got the 12 core anyways and the 16 core I got ended up with me getting and paying for faster ram than I can use doh).

    So I manually memory frequency option back DOWN to 3200Mhz and didn't touch anything else.

    Booted up no drama. Windows shows 3200Mhz no drama. Cpuz and HWinfo get us the following.

    qyUQsWP.png


    H1EAoov.png


    So far no crashes but I haven't done anything to push it. To be fair, I doubt I will notice one iota of difference but I did pay for it and I had been meaning to get it sorted - was just quite anxious about faffing around in the bios just for something that likely won't make my experience any way 'quicker' in a way I will feel. I mean, I went from 2133 15 15 15 36 51 to 3200 18 19 19 39 75 - do I win? Oddly, the DOCP in bios seemed to be aiming for 17 and 19 and I got 18 and 19. Close enough?
    Click to expand...
    The timing are listed in clock cycles, but the chip is really rated in nanoseconds. That means you should scale these with the clock speed.

    If rated for 3600 CL 17, that is 9.44ns, the equivalent timing at 3200 would be CL 16 (10ns) or CL 15 (9.375), so you could reduce those. Generally only frequency and voltage are impacted by the chipset or processor, the rest of the timings are internal to the chip (although a stable voltage is required with DDR4, they had not moved voltage regulation to the module yet).

    Your equivalent 3200 timings to that XMP profile at 3600 would be:
    Rated at 3600
    17 (9.44ns)
    19 (10.55)
    19 (10.55)
    39 (21.66)
    58 (32.22)

    Lowest without exceeding rated timings at 3200
    16 (10ns)
    17 (10.62)
    17 (10.62)
    35 (21.87)
    52 (32.5)

    You could try reducing some of those by one if desired. You can also generally dramatically increase trefi (higher is better) if you have OK cooling (this is the refresh interval, which defaults relatively low as it needs to be at high temperature, but room temperature RAM will hold values for seconds without a refresh, and scaling is not linear such that it is almost always set far lower than is necessary).
     
    Last edited:
    C

    cerberusTI

    Ars Tribunus Angusticlavius
    6,449
    Subscriptor++
  • #23
    It would be a high single digit percentage performance increase at best, it could be 10% if you also look at subtimings and set trefi higher, but is unlikely to exceed this.

    Leaving it is fine (I left the last one I set up at JEDEC timings for weeks due to lack of time to set it), but if you have not done testing yet, you have also not spent much effort yet.

    What I would personally do is find a utility which tells you which memory chips you have, look around for suggested timings for that chip (including subtimings), and set them before you test. If that fails, back off to the rated timings above and slightly increase other subtimings. If that fails, voltage is likely too low to run four modules.
     
    yd

    yd

    Ars Legatus Legionis
    20,764
    Subscriptor++
  • #24
    Correct me if I am wrong but my ram, as now set, would be likely around 50% faster than it was? If we are honest, I can't type 50% faster and I don't really think anything else is doing anything 50% faster. Maybe if I use my 4k convert video to mp3 it would be faster but that is alreay practically instantaneous. I am not sure the 10% (tops) is worth the potential for a blue screen or no boot. Given I theoretically got a 50% boost and nothing seems to have changed, stability and no crashes is kinda where I think, for me, that is solid.
     
    yd

    yd

    Ars Legatus Legionis
    20,764
    Subscriptor++
  • #26
    whoisit said:
    Yeah. Honestly if you are happy with it how it is now, and it's stable, leave it alone. Tightening and testing timings can become like a black hole. Once you cross the event horizon, you're never seen again.
    Click to expand...
    I changed this one number from 18 to 17777777777777777777777777777777777777777777777777777777777 :oops:
     
    • Haha
    Reactions: BigLan
    C

    cerberusTI

    Ars Tribunus Angusticlavius
    6,449
    Subscriptor++
  • #27
    Your first post was at a first word ideal latency (cl) of 14.44ns, your last at 11.25ns, the manufacturer rated it for 9.44ns at 1.35v or 14ns at 1.2v.

    It is not a big performance difference, although I would still make sure that voltage is adequate as while that 3200 rating is due to the JEDEC standard, you are outside JEDEC timings. An extended test is the usual way to be sure the current settings work, and you should still do so.

    I usually do not try extremely tight sub timings, and back off from the best passing setings a little bit before retesting as well (I like a little bit of margin).
     
    yd

    yd

    Ars Legatus Legionis
    20,764
    Subscriptor++
  • #28
    cerberusTI said:
    Your first post was at a first word ideal latency (cl) of 14.44ns, your last at 11.25ns, the manufacturer rated it for 9.44ns at 1.35v or 14ns at 1.2v.

    It is not a big performance difference, although I would still make sure that voltage is adequate as while that 3200 rating is due to the JEDEC standard, you are outside JEDEC timings. An extended test is the usual way to be sure the current settings work, and you should still do so.

    I usually do not try extremely tight sub timings, and back off from the best passing setings a little bit before retesting as well (I like a little bit of margin).
    Click to expand...
    I am trying to remember, which is the ram stress tester of choice....I probably have it somewhere. I can see if it makes these settings crap out. It certainly isn't impacted surfing the net watching/watching/videos or playing basic games - I haven't played PUBG which is the only fps I pay to see if that craps out but I know it gets the load screen no drama.
     
    continuum

    continuum

    Ars Legatus Legionis
    94,897
    Moderator
  • #29
    Memtest86 or Memtest86+, I forget which is currently the most current of the two, but so long as your motherboard/CPU is supported either should be fine.

    I usually hit with a few runs of Linpack (via IntelBurnTest or the newer Linpack Xtreme) with appropriate memory settings to max out and test all of the installed memory (well, as much as realistically possible). Prime95 too.

    I've had systems pass multiple stress tests for 24 hours plus each test then fail a game or something else so yeah, all depends how careful you want to be. My personal experience with 5950X's and 16GB modules however is limited, but my luck with multiple 5950X's and 4x32GB modules is that for long-term stability, DDR4-3200 CL16 1.35v (depending on module) is pretty achieveable, I have not had good luck with DDR4-3600 at any voltage or latency combination.

    That said you are running 16GB modules which is a completely different beast.
     
    • Like
    Reactions: cerberusTI
    C

    cerberusTI

    Ars Tribunus Angusticlavius
    6,449
    Subscriptor++
  • #30
    The tests listed above are great for finding memory issues, although there is a major factor they do not test, which is that DRAM is temperature sensitive, and sensitive to voltage swings on DDR4 as well (DDR5 moves this regulation to the module).

    A long test is less useful in some ways than a hot test, and something like https://www.ocbase.com/ offers good options to get an idea if your system will be stable under load. Linpack does do this to some degree (that on its own is a good general test of CPU and memory settings), but maxing everything out in a power test including the video card can also show issues you could see with gaming (or in my case, the development of bulk AI indexing calculations so the later task will run faster).

    It is not strictly a substitute, as it is does not check all memory. It can check memory you are not currently using for the OS or other applications, but it can check it under more adverse conditions.

    I cannot really be specific either, my desktop DDR4 system was an 11900K with DDR4 4000 at 9ns, but it is two 8GB modules. I had it set a fair bit faster than that (with no issues), but set it back to XMP default settings and much lower power limits for the CPU as it became a file server. That is also a completely different beast.

    I will note that a 7800X3D I have passed memory testing at 6400 with fairly aggressive settings under basically everything, until all components were maxed out for quite a while at once in that test, but eventually many memory errors presented themselves in a short time, and it was not holding values well at all. Would that ever be an issue in normal use? I do not know, but I would prefer it be stable under even very hot conditions, so I do like a test which tests all components at once.

    There is also, as some have noted, the safety and assurance of JEDEC timings.
     
    Last edited:
    yd

    yd

    Ars Legatus Legionis
    20,764
    Subscriptor++
  • #31
    Am I really not within Jedec specs - this is the ram and it shows 17 19 19 39 and rated for 3600

    735ny2c.png


    I am at 18 19 19 39 which would be 'easier' timings, not pushing it out of bounds.

    I really don't recall this ram stuff being this tricky on the Intel boards, that said, maybe I should look and see if the xmp thing didn't make the latency terrible; however, for an htpc it really doesn't matter and my wife doesn't care on her desktop so maybe ignorance is bliss for those.

    edit - added the ram as bought from NewEgg back in the day
     
    Last edited:
    C

    cerberusTI

    Ars Tribunus Angusticlavius
    6,449
    Subscriptor++
  • #32
    The fastest DDR4 standard is 3200 20-20 20 at 1.2v. I do not see that in the SPD for your module though, it may want 1.35 even at that speed.

    Anything at a higher frequency, lower timings at that frequency, or an increased voltage, is overclocking (and should be tested).

    On Intel systems you generally just need to enable XMP, but that is their overclocking spec, and even on Intel that would only be very likely to work if you only had two modules, otherwise you may still be reducing it manually as you are here. DOCP may just work for you as well if you only fill two of the slots, as they intended unless you bought that as a single four pack of modules.
     
    Last edited:
    • Like
    Reactions: continuum
    yd

    yd

    Ars Legatus Legionis
    20,764
    Subscriptor++
  • #33
    Four modules baby - why get 32 gigs of ram when you can have 64! Admittedly, I never use more than about 25 just kinda eyeballing things but moar is always moar betterer!

    If I move any of those timings down to say 20 20 20 or even 19 20 20 I am still 'out of spec' but the more I move those numbers down the more loss I experience correct? As mentioned, my goal was to get a nice bump off the 2133Mhz I was jammed with by it not posting back when I did the build. It turns out that I don't think there was a bios problem at the time but rather I was trying to push it to 3600Mhz which was not cool with the processor.

    At this juncture, I have not crashed with them as shown 17 19 19 19 39 and 3200Mhz which has to be an improvement over the 2133Mhz settings.
     
    C

    cerberusTI

    Ars Tribunus Angusticlavius
    6,449
    Subscriptor++
  • #34
    yd said:
    Four modules baby - why get 32 gigs of ram when you can have 64! Admittedly, I never use more than about 25 just kinda eyeballing things but moar is always moar betterer!

    If I move any of those timings down to say 20 20 20 or even 19 20 20 I am still 'out of spec' but the more I move those numbers down the more loss I experience correct? As mentioned, my goal was to get a nice bump off the 2133Mhz I was jammed with by it not posting back when I did the build. It turns out that I don't think there was a bios problem at the time but rather I was trying to push it to 3600Mhz which was not cool with the processor.

    At this juncture, I have not crashed with them as shown 17 19 19 19 39 and 3200Mhz which has to be an improvement over the 2133Mhz settings.
    Click to expand...
    You are out of spec, but the spec is very conservative, and you bought memory a vendor has theoretically cleared at a higher speed than any specification (what that means is basically up to them, but that is one of the better vendors, and it should be able to do what they say).

    Frequency also requires the processor be able to do this (it can likely do 3600, even if they only guarantee 3200), and how many modules you are using (likely why it did not work). It is much harder to get the CPU to successfully run four modules at a high speed, and relaxing frequency is often required if you use four. You could possibly tell the memory controller to run at 2:1, which would allow this to work (Intel does that, which is why it takes higher speeds), but it would add a little bit of latency, and is probably not a gain for gaming for that reason.

    Timings are usually internal to the memory chip, but the chip does care about stable voltage, and likely will only do this reliably if provided the 1.35v it asks for (which is not within any JEDEC spec on voltage or timings, but the vendor has binned it such that it should do that anyway). It is much more likely you can set what is rated for timings, even with four modules. If not, it is usually possible to very slightly increase voltage to compensate for the extra electrical load.

    Your current timings and transfer rate are much better than your initial settings, but you should still test it, as it is not within spec.
     
    • Like
    Reactions: continuum
    yd

    yd

    Ars Legatus Legionis
    20,764
    Subscriptor++
  • #35
    I have to say I am a bit pissed at myself for not checking that a chip also had to match ram speed - I thought ram speed was mobo support related 'only' so bought the chip on the basis of tons of cores and the ram on the basis of at the time 3600 was 'fast'.

    I definitely could have saved about 150 bucks and did the 12 core AMD option (5900?) at the time because nothing ever uses 16 cores so I might as well have had the higher base clock and boost speed. That said, this 16 core version ticks along at 4.3GHz on average
     
    C

    cerberusTI

    Ars Tribunus Angusticlavius
    6,449
    Subscriptor++
  • #36
    It is not just the CPU, which can very likely do that under the same conditions the modules and motherboard can, they just do not want to take an RMA over it. In some ways that is more honest, as the rate AMD tends to give as the max is what you can reasonably expect four modules at top end JEDEC timings to do.

    Everyone else does a poor job of explaining that you cannot usually populate all four slots and get those speeds, and it is a common issue. You can likely set the better wall time timings and have it work though, and that is more meaningful in gaming and most other workloads than the signaling rate. It just has difficulty signaling from the CPU to the modules at that rate with four modules due to reflection issues unless you get lucky in a few ways.
     
    Last edited:
    • Like
    Reactions: continuum
    continuum

    continuum

    Ars Legatus Legionis
    94,897
    Moderator
  • #37
    yd said:
    At this juncture, I have not crashed with them as shown 17 19 19 19 39 and 3200Mhz which has to be an improvement over the 2133Mhz settings.
    Click to expand...
    yd said:
    this is the ram and it shows 17 19 19 39 and rated for 3600
    Click to expand...
    Just leave it, you'll be fine. It's overclocked but as noted and discussed, your 5950X should be fine at those settings.

    cerberusTI said:
    Frequency also requires the processor be able to do this (it can likely do 3600, even if they only guarantee 3200), and how many modules you are using (likely why it did not work). It is much harder to get the CPU to successfully run four modules at a high speed, and relaxing frequency is often required if you use four.
    Click to expand...
    ++;
     
    yd

    yd

    Ars Legatus Legionis
    20,764
    Subscriptor++
  • #38
    Just for peace of mind, I ran a 32 instance Prime95 torture test for a simplicity for a few hours. 100% cpu usage of the 16 core 5650x and 100% usage of the 64gb ram at 3200Mhz with those timings above. No drama. CPU temp sat at around 72c in a 25c room.

    For chuckles, I checked the htpc and wife's desktop and both of those I used xmp for when I set them up and both of those are running at 3600Mhz with 'normal' aka good timings.

    Seems the key mistake on my part was believing this setup could run ram at 3600Mhz not knowing that ram timings were tied to chips. I probably should have went with the 12 core at the time but I didn't so whatevs - I never see 100% usage on this cpu, not even close. Pathetic that more things are not able to use cores more effectively/completely.
     
    yd

    yd

    Ars Legatus Legionis
    20,764
    Subscriptor++
  • #40
    Indeed. At the time, the difference was like, 200 bucks maybe. My thought was, who knows so why not have 16 cores! Turns out 16 cores are under utilized as frankly would be a 12 cores. Performance wise, I would wager (a bit) they are identical performance in my usage - the difference is I have 4 more cores (which are not usually used) and the base clock is 3.4Ghz vs the 12 core being 3.7Ghz. So theoretically the 12 would run a bit quicker.....however, their boost clocks matter too and the 12 core boosts to 4.8Ghz while the 16 boosts to 4.9Ghz. And based on my watching taskmanager, the 5950x routinely doesn't even max out the boost clock running around 4.3Ghz so the 5900x wouldn't max out either. same performance, so 4 more cores on tap 'just in case'. Good deal!

    I have now successfully justified the extra cores haha!
     
    You must log in or register to reply here.
    Jump to latest Follow Reply
    Jump to latest Follow Reply