What memory speeds can you expect from Ryzen?

Discussion in 'AMD x86 CPUs and chipsets' started by Sgt Bilko, Jan 17, 2019.

  1. Sgt Bilko

    Sgt Bilko Member

    Joined:
    Mar 5, 2015
    Messages:
    984
    Location:
    Canberra(ish)
    So as the title says, what can you expect?

    Well in an attempt to help out present and future Ryzen users I decided to gather what DDR4 memory I have here and test for daily stability without going too in depth (only changing voltage and primary timings).

    My test system will be made up of:

    CPU: Ryzen 7 2700X
    CPU Cooler: Noctua NH-U9S
    Motherboard: X470 AORUS Gaming 7 WiFi (F4 BIOS, AGESA 1.0.0.4)
    Memory: Varied
    Graphics Card: RX 480
    PSU: Corsair HX1200i
    OS: Windows 10 64 bit

    The CPU was locked at a static 4.0Ghz for consistency, for the majority of my testing the SOC voltage was running at 1.1v set in BIOS which is closer to 1.08-1.09v real but some kits appeared to prefer a higher SOC voltage once manually tuned.

    As a personal rule of thumb I consider 1.4v Dram to be safe for daily operation on DDR4 but of course not all memory ICs react the same way to voltage and degradation so for testing purposes I'm ok with pushing the boundary a little but I don't recommend over 1.4v Dram for a daily setup.

    For testing the memory I'm using the Karhu Software RAM Test, I've found this to be an excellent program in testing for instability related to memory and it appears to detect instabilities fairly quickly compared to others I've used in the past but of course with any stability testing your mileage may vary.

    The text below is taken from Karhu Softwares FAQ section in regards to how long the stress test should be run for and coverage rate vs error detection.
    I've also included AIDA64 results in addition to the stress test for an indication of performance gain or loss while using various kits at different speeds/timings as well.

    So with most of that out of the way these are the memory kits that I've tested:

    2x8GB Team Group Dark Pro 3466 c16 (Single Sided Samsung B Die ICs)
    2x8GB Team Group Nighthawk RGB 3200 c16 (Single Sided Samsung B Die ICs)
    2x8GB Corsair Vengeance LED 3200 c16 v4.24 (Dual Sided Samsung E Die ICs)
    2x8GB Team Group T-Force Dark 3000 c16 (Single Sided Hynix AFR ICs)
    2x4GB Zadak511 Shield 3000 c16 (Single Sided Hynix MFR ICs)
    4x8GB Team Group Nighthawk RGB 3200 c16 (Single Sided Samsung B Die ICs)

    And that is the sum total of all the DDR4 memory that I happen to have on hand, While it's not alot in the grand scheme I do hope that this will at least help someone :)

    Onto the testing!

    So we'll start off with the 3466 c16 Dark Pro kit from Team Group but at JEDEC speed and timings
    2400Mhz 16-16-16-39-55 with 1.2v Dram and 1.09v SOC, No stability test here as I feel as though it's not needed.

    [​IMG]

    Next was trying XMP which in this case is 3466Mhz 16-18-18-38-56 at 1.35v Dram
    This one I tested for an hour coverage which I'm happy to say performed flawlessly.
    [​IMG]

    I decided I'd try and push my luck and for for 3600Mhz at the same XMP timings and voltage
    Sadly this one did fail but I will attempt to revisit it before too long.
    [​IMG]

    Onto the next kit!

    Some more Samsung B Die but this time with an RGB flavour in the form of Team Groups Nighthawk RGB
    The XMP for this kit is 3200Mhz 16-18-18-38-56 with 1.35v Dram and as expected was stable for me with those settings.
    [​IMG]

    Next was to bump up the speed and voltage (1.4v Dram) while lowering the timings a little.
    3466Mhz 16-16-16-36-48 with ProcODT set to 60 ohms.
    [​IMG]

    Instead of pushing for higher speed I then decided to lower the speed and try and tighten the primary timings
    So I ended up with 3200Mhz 14-14-14-34-48 with 1.4v Dram and ProcODT again set to 60 ohms.
    [​IMG]

    As a final test with this configuration I stuck with the same speed as before but again lowered the timings a little but this required a touch more Dram voltage to be stable so now we have 3200Mhz 14-14-14-28-36 with 1.42v Dram and again ProcODT set to 60 ohms.
    [​IMG]

    Next up is my only Dual Sided memory kit, Now this kit of Corsair Vengeance LED predates Ryzen and it being Dual Sided I was not expecting great results from it but we'll see what happens.

    First up is the JEDEC spec which here is 2133Mhz 15-15-15-36-51 at 1.2v Dram
    No stability test as again I don't believe it to be needed
    [​IMG]

    This is the only kit I tested that failed to run stable at its XMP which doesn't surprise me but that doesn't mean we can't work with it a little, While not stable at the XMP speed of 3200 it was in fact stable at 3133Mhz with 1.4v Dram and the XMP timings.
    [​IMG]

    To get that small bump up to 3200Mhz this kit needed 1.45v Dram to become stable which is above what I'd recommend for daily settings.
    [​IMG]

    Obviously we couldn't really be going up in speed or down in timings next was to drop the memory speed and timings to see if more could be gained there so at 1.38v Dram it managed 3000Mhz 16-18-18-36-54
    [​IMG]

    And onto another kit!
    Next is some Single Sided Hynix AFR which here is in Team Groups T-Force Dark 3000 kit.
    So we'll start of at the XMP of 3000Mhz 16-18-18-38-56 which is nice and stable
    [​IMG]

    Next we'll bump up the speed to 3200Mhz, change dram voltage to 1.4v, and keep the timings where they are.
    [​IMG]

    Since 3200 ran so smoothly I decided to push a little higher and go for 3266Mhz 16-18-18-40-60 with 1.41v Dram and 1.15v SOC and which did fail but it made it over half way through the test as where most failure occurred within the first 5-10 minutes.
    [​IMG]

    And finally onto the last memory type I own, Hynix MFR which here is in a couple of Zadak511 Shield RGB sticks.

    Starting off with the XMP of 3000Mhz 16-18-18-38-57 which was stable
    [​IMG]

    Next was to up the voltage a little to 1.4v Dram and lower the timings down so we ended up with 3000Mhz 16-16-16-34-50
    [​IMG]

    But that passed pretty easy so bumping the dram voltage a tiny amount again to 1.41v and SOC to 1.1v resulted in a pretty respectable 3000Mhz 14-16-16-28-48
    [​IMG]

    And that was where I stopped for 3000Mhz, for 3200Mhz I stepped the voltages back down to 1.35v Dram and 1.1v SOC so we ended up with 3200Mhz 16-18-18-38-57 stable
    [​IMG]

    Moving on from there I lowered a couple of timings and reran which resulted in 3200Mhz 16-16-16-38-57 stable
    [​IMG]

    and finally bumping up the dram voltage to 1.42v and the SOC to 1.125v resulted in 3200Mhz 14-16-16-34-50 stable, pretty good result imo
    [​IMG]]

    And finally the 4x8GB NightHawk RGB results (4 x Single Sided B Die), Since this is more memory I needed to do a longer stress test which is part of the reason why these took longer.

    Starting off with the XMP of 3200 16-18-18-38-56 at 1.35v Dram, all nice and stable
    [​IMG]

    From there I bumped Dram voltage to 1.4v, SOC to 1.125v and went for 3200 14-14-14-34-51 which was stable
    [​IMG]

    From there I went up in speed and timings but leaving Dram voltage at 1.4v, SOC was dropped back to 1.1v, and we got 3466Mhz 16-18-18-38-56 Stable
    [​IMG]

    And finally, 3600Mhz on 4x8GB DIMMs, this one took a while to test and I probably spent the better part of a day just on this configuration alone, I'm including it so others could try it if they wanted but I never could get it stable enough for me, my best result I manually stopped at 20mins.
    3600Mhz 16-18-18-38-56 with 1.45v Dram, 1.09v SOC and 48ohm ProcODT
    [​IMG]

    And onto Threadripper!

    System specs for this one are:

    CPU: Ryzen Threadripper 1900X/2920X
    CPU Cooler: Noctua NH-U14S TR4
    Motherboard: X399 AORUS Gaming 7 (F11e BIOS, AGESA 1.1.0.1)
    Memory: 4x8GB TeamGroup Nighthawk RGB 3200Mhz c16
    Graphics Card: RX 480
    PSU: Corsair HX1200i
    OS: Windows 10 64 bit

    So with this I tested both 1st Gen and 2nd Gen Threadripper seeing as I was only using one memory kit.
    Both CPUs were again set to 4.0Ghz in an effort to try and keep everything as consistent as possible, all testing was done with the memory in Distributed mode (Set in BIOS) as that's the default config.
    RTC doesn't read SOC voltage or ProcODT for me here so unless I say otherwise it's set to 1.1v and 53.3 respectively :)
    First up let's start with the 1900X!

    Straight out of the box default 2133Mhz
    [​IMG]

    Then of course onto XMP (3200Mhz 16-18-18-38-56) which as expected was stable
    [​IMG]

    Next was to drop the timings so here is 3200Mhz 14-14-14-36-75 at 1.38v Dram and 1.15v SOC
    [​IMG]

    From there I took the leap straight to 3600Mhz 16-18-18-38-84 with 1.4v Dram, 60ohm ProcODT and 1.125v SOC
    [​IMG]

    There wasn't anything I could gain from this point so I moved onto the 2920X, at the request of glnn_23 I've added some NUMA results alongside the UMA ones, these are using a different OS as the initial one I was using for the tests had been corrupted after more than a few blue screens and generally being unstable but they are using the same settings as their UMA counterparts.

    As before we start off with 2133Mhz at default to get a baseline
    [​IMG]
    No NUMA result for 2133Mhz

    And then move onto the XMP of 3200Mhz 16-18-18-38-56 which was stable
    [​IMG]

    [​IMG]

    After that I tightened up the timings a little to get 3200Mhz 14-14-14-34-48 with 1.36v Dram, 1.1v SOC and 48ohm ProcODT
    [​IMG]

    [​IMG]

    From there I reset XMP but changed the speed to 3466Mhz and set Dram to 1.4v with ProcODT at 53.3ohms
    [​IMG]

    [​IMG]

    And finally 3600Mhz, this one was a bit of a pain and I suspect my OS had started to give up at this point but while it passed the 1 hour stability test (it got a single error just after the hour mark) it took voltages I'm not personally comfortable with on a daily system.

    3600Mhz 16-18-18-38-56 with 1.45v Dram, 1.1v SOC and 60ohm ProODT
    [​IMG]

    [​IMG]


    Either way I hope this helps someone out and if there's anything you'd like to know please ask and I'll do my best to answer it :D

    That's all folks!
     
    Last edited: Jan 24, 2019
  2. darkbastard

    darkbastard Member

    Joined:
    Sep 17, 2004
    Messages:
    592
    Location:
    Duranbah - near Kingsclif
    Nice Work Sarge!!

    Very interested in your 4*8GB BDie tests as I am running 4 sticks of Gskill C14 Bdie myself on a CH 6.

    My 24/7 settings are 14-14-14-28-50 1T but I should probably push it a bit and see what I can get stable at. I know I have run it at C12 per the stilts fast 3200 2 *8 settings and it was fine for weeks. But that doesnt make it stable :)
     
  3. OP
    OP
    Sgt Bilko

    Sgt Bilko Member

    Joined:
    Mar 5, 2015
    Messages:
    984
    Location:
    Canberra(ish)
    Thanks mate, Not sure I'll be pushing as low as c12 but we'll see how it goes :)
     
  4. glnn_23

    glnn_23 Member

    Joined:
    Jan 19, 2012
    Messages:
    442
    Great work Sgt Bilko

    I've done a little memory testing with a 2700x. I found the biggest difference in what I can get out of the memory depends on the choice of motherboard and the amount of ram sticks used.

    I tried the Asus Crosshair VII and the Asus Strix itx.

    The itx ,only having 2 dim slots, easily performed the best. Ran as high as 3950c18 running A64 benchmark
    and pass GSAT at 3700c14
    The Crosshair couldn't do this with 2 dimms and running 32Gb with 4 dimms was very difficult.

    Here's the settings I used for 3700c14 1.425v. SOC 1.1v. G.Skill TridentZ 4266
    https://www.overclock.net/forum/attachment.php?attachmentid=228928&d=1541231359
     
    Ratzz and darkbastard like this.
  5. OP
    OP
    Sgt Bilko

    Sgt Bilko Member

    Joined:
    Mar 5, 2015
    Messages:
    984
    Location:
    Canberra(ish)
    Cheers mate, Motherboard does make a difference and DIMM amount certainly does make a difference as well for sure.
    That's some impressive results you've got there, well done :D

    Also, updated OP with the 4x8GB testing on X470 and X399
     
  6. glnn_23

    glnn_23 Member

    Joined:
    Jan 19, 2012
    Messages:
    442
    Sgt Bilko have you tried running with the memory interleave set to channel and die or NUMA / UMA ?
     
  7. OP
    OP
    Sgt Bilko

    Sgt Bilko Member

    Joined:
    Mar 5, 2015
    Messages:
    984
    Location:
    Canberra(ish)
    Forgot to add it to the OP, all X399 testing was done with the memory in Distributed mode (Set in BIOS) as that's the default config.
     
  8. glnn_23

    glnn_23 Member

    Joined:
    Jan 19, 2012
    Messages:
    442
    Can you do a run in NUMA with the same settings to compare to Distributed ? I would expect a big improvement in Aida64 in Read and Latency.
     
  9. OP
    OP
    Sgt Bilko

    Sgt Bilko Member

    Joined:
    Mar 5, 2015
    Messages:
    984
    Location:
    Canberra(ish)
    I'll see about it, as I suspected earlier the OS is toast.
    Need to completely reinstall Windows so it won't be today
     
  10. OP
    OP
    Sgt Bilko

    Sgt Bilko Member

    Joined:
    Mar 5, 2015
    Messages:
    984
    Location:
    Canberra(ish)
    glnn_23 Updated OP with NUMA results for 3200 c16, 3200 c14, 3466 c16 and 3600 c16
    They are on a different windows install but after doing some comparison runs with UMA and comparing to the ones in the OP it's margin of error.
     
    darkbastard likes this.
  11. glnn_23

    glnn_23 Member

    Joined:
    Jan 19, 2012
    Messages:
    442
    Thanks again Sgt Bilko for the great work. Quite a performance gain there in the Aida64 benchmark switching to NUMA.

    Having used the x399 Creation MB I wonder if something like the Asrock X399M only having 4 dim slots would be able to push memory further .
     
  12. MUTMAN

    MUTMAN Member

    Joined:
    Jun 27, 2001
    Messages:
    6,361
    Location:
    4109
    are you planning on a repeat of these tests (and add faster sticks too) with the 3000/x570 combos (once released)
    looking forward to seeing bigger better faster numbers :)
     
  13. OP
    OP
    Sgt Bilko

    Sgt Bilko Member

    Joined:
    Mar 5, 2015
    Messages:
    984
    Location:
    Canberra(ish)
    Maybe!

    We'll see what happens :)
     
    MUTMAN likes this.
  14. Hotrod2go

    Hotrod2go Member

    Joined:
    Jun 17, 2018
    Messages:
    276
    Location:
    Tas
    Interesting thread, Be curious to see testing done on B450 platform? :)
    After all, some of us gamers don't need more than 1 monitor so multiple PCIe slots are useless.
     
  15. OP
    OP
    Sgt Bilko

    Sgt Bilko Member

    Joined:
    Mar 5, 2015
    Messages:
    984
    Location:
    Canberra(ish)
    It's something I can certainly look at adding in when I've got some spare time :)
     
    Hotrod2go likes this.
  16. the_antipop

    the_antipop Member

    Joined:
    Jun 11, 2005
    Messages:
    1,792
    Location:
    Gold Coast
    Just had a question about Ryzen memory.
    At the moment I have 2x 8GB sticks in my rig running in dual channel. If I then added another 2x 8GB sticks of the same memory, would it then go back to single channel? Or would I still be able to run in dual channel? Specs in sig.
     
  17. OP
    OP
    Sgt Bilko

    Sgt Bilko Member

    Joined:
    Mar 5, 2015
    Messages:
    984
    Location:
    Canberra(ish)
    So all up running 4x8GB?
    If so it'd still be Dual Channel but you'd be running 2DPC (2 DIMMs Per Channel) opposed to 1DPC like you are now :)
     
    the_antipop likes this.
  18. shane41

    shane41 Member

    Joined:
    Nov 10, 2008
    Messages:
    6,201
    Location:
    dice que no sabe
    Done very well on your testing Bilko :thumbup: I did stumble across the thread & I bookmarked it for a read tomorrow.
    Big day & I'm fucked
     
  19. Etherate

    Etherate Member

    Joined:
    Dec 16, 2003
    Messages:
    931
    Location:
    SA
    thanks bilko, very handy
     
  20. Ratzz

    Ratzz Member

    Joined:
    Mar 13, 2013
    Messages:
    8,611
    Location:
    Cheltenham East 3192
    Great thread bilko, very comprehensive. I can add one more set of numbers if that helps anyone. I'm using Hyper-X Predator c19-4000 (2x8 B-Die) on an X470 Taichi with a 2700X. I've tried multiple settings, I can boot at c19-3733, but staying in Windows for more than 30 seconds is not possible. Just loading a couple of instances of CPU-Z is enough to lock up.

    I can however run c14-3600 1T stable, in fact it's my 24/7. I'm running ram @1.35v.

    [​IMG]

    [​IMG]


    Not sure if I've gotten lucky with a pair of strong sticks, or its the brand/model, or my board, or my awesome skillzzz :lol: but I'm not seeing many(any?) people run c14-3600 on a 2000 CPU. I had to mess around a fair bit with sub timings to get those numbers stable though, its a lot more than just primary timings.

    I've tried dropping latency but I can't do c12 unless I drop frequency back to 2933.

    I've never tried 4 sticks, I have yet to get around to getting a second set of these. I'm hopeful that since my Taichi has T-Topology rather than daisy chain, I'll be able to get the same c14-3600 with all dimms populated, unless I luck out and get a weak stick. When I get a Round Tuit, I'll update on that.

    I'm hopeful that when I buy a Zen2 later in the year I can get better than the current numbers too. I see that glnn_23 is getting 3733 stable (I assume? at least he is running benchies at that speed, that's further than I can get) from a 3900X so fingers crossed.
     
    Last edited: Jul 14, 2019

Share This Page

Advertisement: