-
Welcome to Overclockers Forums! Join us to reply in threads, receive reduced ads, and to customize your site experience!
3900x RAM & memory controller OC
- Thread starter lightingman117
- Start date
I don't get memory comparison tests in Cinebench R20 which barely reacts to memory/IF changes.
In my tests Cinebench results at various memory and IF frequencies give back +/-100 point results (when there was already 7k+). This is between 3200, 3600, ..., 4266+. While switching to 1:2 IF ratio, there is not so big but still visible performance drop. Something like 3600 CL16 at 1:1 can be compared to 4266 CL18 1:2 using 2x16GB sticks. Simply memory performance doesn't affect this benchmark enough to say that something runs wrong, especially when a typical error margin is about 50 points (rerun the benchmark multiple times and you will see it).
Cinebench scores can be different because of many factors. Some examples which you can check on your own. Set the CPU at auto clock + 1.35V voltage ( or -0.01V offset ) and in the second case leave everything at auto. On the 3900X you will see up to ~200 points difference in Cine R20. Another case, run everything at auto with better and worse cooler (let's say +/- 15°C) and you will see ~200 points difference. One more, set power saving in power plan in Windows vs high performance and the same, 200-300 points difference.
In my experience, the only case when ASRock or ASUS motherboard couldn't run at 1:1 with RAM at 3600 was when the motherboard had problems at the memory training stage. Then it was forcing lower IF. However, it's not a motherboard issue but normal behavior when memory or the IMC can't handle some settings. I had that almost only when I was setting everything manually.
In my tests Cinebench results at various memory and IF frequencies give back +/-100 point results (when there was already 7k+). This is between 3200, 3600, ..., 4266+. While switching to 1:2 IF ratio, there is not so big but still visible performance drop. Something like 3600 CL16 at 1:1 can be compared to 4266 CL18 1:2 using 2x16GB sticks. Simply memory performance doesn't affect this benchmark enough to say that something runs wrong, especially when a typical error margin is about 50 points (rerun the benchmark multiple times and you will see it).
Cinebench scores can be different because of many factors. Some examples which you can check on your own. Set the CPU at auto clock + 1.35V voltage ( or -0.01V offset ) and in the second case leave everything at auto. On the 3900X you will see up to ~200 points difference in Cine R20. Another case, run everything at auto with better and worse cooler (let's say +/- 15°C) and you will see ~200 points difference. One more, set power saving in power plan in Windows vs high performance and the same, 200-300 points difference.
In my experience, the only case when ASRock or ASUS motherboard couldn't run at 1:1 with RAM at 3600 was when the motherboard had problems at the memory training stage. Then it was forcing lower IF. However, it's not a motherboard issue but normal behavior when memory or the IMC can't handle some settings. I had that almost only when I was setting everything manually.
Last edited:
OP
- Joined
- Apr 27, 2020
- Thread Starter
- #43
Definitely agree that Cinebench gives fluctuations and with Ryzen is especially dependent on CPU temp.
I usually run 3 runs and average the result if I care about real numbers; if I'm just testing it quick to see the difference, I take that into account for not expected numbers (see below first result).
I am having memory training issues at 3800/1900, it's stable once booted. But sometimes it doesn't post right away, a bit more memory voltage smoothed things out I think (1.384v & 36.3C [43C max P95 MemStress]). I might run slower than 3800/1900 for my daily. But we'll see how this goes for now.
& good point about R20 not stressing memory components much, but I have moar data now
----------------------------------------------------------------------
3200/1600 MCLK/FCLK
Stock core clock speed (4GHz AC)
--------------------------------
IBT 134 GFlops
R20 7213 (cold),7163,7170
read 49630
write 47379
copy 52123
latency 75.8
--------------------------------
3600/1800 MCLK/FCLK
Stock core clock speed (4GHz AC)
--------------------------------
IBT 140 GFlops
R20 7182
read 55986
write 53468
copy 58625
latency 69.6ns
--------------------------------
3333/1900 MCLK/FCLK
Stock core clock speed (4GHz AC)
--------------------------------
IBT 140 GFlops
R20 7147,7132,7155
read 51126
write 49120
copy 52690
latency 86.7ns
--------------------------------
3800/1900 MCLK/FCLK
Stock core clock speed (4GHz AC)
--------------------------------
IBT 140 GFlops
R20 7165,7155,7152
read 58762
write 55968
copy 61745
latency 67.0ns
--------------------------------
3800/1900 MCLK/FCLK
4.3GHz @ 1.34v
--------------------------------
IBT 155 GFlops
R20 7630,7686,7680
read 58369
write 56427
copy 61488
latency 66.9
I usually run 3 runs and average the result if I care about real numbers; if I'm just testing it quick to see the difference, I take that into account for not expected numbers (see below first result).
I am having memory training issues at 3800/1900, it's stable once booted. But sometimes it doesn't post right away, a bit more memory voltage smoothed things out I think (1.384v & 36.3C [43C max P95 MemStress]). I might run slower than 3800/1900 for my daily. But we'll see how this goes for now.
& good point about R20 not stressing memory components much, but I have moar data now
----------------------------------------------------------------------
3200/1600 MCLK/FCLK
Stock core clock speed (4GHz AC)
--------------------------------
IBT 134 GFlops
R20 7213 (cold),7163,7170
read 49630
write 47379
copy 52123
latency 75.8
--------------------------------
3600/1800 MCLK/FCLK
Stock core clock speed (4GHz AC)
--------------------------------
IBT 140 GFlops
R20 7182
read 55986
write 53468
copy 58625
latency 69.6ns
--------------------------------
3333/1900 MCLK/FCLK
Stock core clock speed (4GHz AC)
--------------------------------
IBT 140 GFlops
R20 7147,7132,7155
read 51126
write 49120
copy 52690
latency 86.7ns
--------------------------------
3800/1900 MCLK/FCLK
Stock core clock speed (4GHz AC)
--------------------------------
IBT 140 GFlops
R20 7165,7155,7152
read 58762
write 55968
copy 61745
latency 67.0ns
--------------------------------
3800/1900 MCLK/FCLK
4.3GHz @ 1.34v
--------------------------------
IBT 155 GFlops
R20 7630,7686,7680
read 58369
write 56427
copy 61488
latency 66.9
Last edited:
OP
- Joined
- Apr 27, 2020
- Thread Starter
- #45
Thanks for the help! I promise I had done a fresh start previously... I wouldn't have bothered y'all. I guess the magic was in the reset button & letting the memory train a few times.
Thanks again!
Not yet, I tried ASRock's utility from software, works...okay. Except it holds CPU freq at whatever you set it to. Useful for rendering or benchmarks...not so much for daily OC.
I tried BIOS OC too, but it's similar.
RM really is nice IMO. & it works well for CPU OC; not great for mem/FCLK tho.
BTW, according to pg.11 the RM guide says it isn't supposed to mess with BIOS after reboot. But who knows. https://www.amd.com/system/files/documents/ryzen-master-quick-reference-guide.pdf
Thanks again!
Not yet, I tried ASRock's utility from software, works...okay. Except it holds CPU freq at whatever you set it to. Useful for rendering or benchmarks...not so much for daily OC.
I tried BIOS OC too, but it's similar.
RM really is nice IMO. & it works well for CPU OC; not great for mem/FCLK tho.
BTW, according to pg.11 the RM guide says it isn't supposed to mess with BIOS after reboot. But who knows. https://www.amd.com/system/files/documents/ryzen-master-quick-reference-guide.pdf
- Joined
- Dec 19, 2012
I was just speaking from my own experience and what I had to do to get things back to normal. It may not "change" values but it is working low level on the motherboard. It also could have been the Windows crash that messed up RM a bit
OP
- Joined
- Apr 27, 2020
- Thread Starter
- #48
Okay so, I've been running this for a few days now. Fully installed all my programs, games, etc. [benchmarks were with vanilla win10 + debloat].
3800/1900 is mostly stable [good for benchmarks], but p95 for a few hours found some errors.
So for my daily, I dropped it to 3733/1866 [MCLK/FCLK] which is super stable once running.
The only issue is every time it reboots I have to re-train the memory. It's getting annoying. [And not easy for the wife to use...which this needs to be]
I'm thinking of returning the ram and getting something that is on the QVL.
I still need 32GB. [which I guess is hard to push for the memory controller].
I don't want to run 4x 8GB sticks because I want the option to go to 64GB in the future (I plan on running this for a long time like my last system (8+years)).
I'd also like to keep running 3733/1866.
So I'm thinking QVL 3600-4000 16GB sticks are best for my application?
https://www.asrock.com/mb/AMD/X570 Taichi/#MemoryMS
Current RAM:
$190 - Newegg - G.Skill Trident Z Neo F4-3600C16D-32GTZNC
Think I have it narrowed down:
$285 - Amazon - 32GB - 3600 - C17 - HX436C17PB3K2/32
$347 - Newegg - 64GB - 3600 - C17 - F4-3600C17Q-64GTZSW
$280 - Newegg - 32GB - 3600 - C16 - CMK32GX4M2K3600C16
$425 - Newegg - 32GB - 3800 - C19 - CMK128GX4M8X3800C19
I'm leaning toward the CORSAIR Vengeance LPX.
What would you get and why?
3800/1900 is mostly stable [good for benchmarks], but p95 for a few hours found some errors.
So for my daily, I dropped it to 3733/1866 [MCLK/FCLK] which is super stable once running.
The only issue is every time it reboots I have to re-train the memory. It's getting annoying. [And not easy for the wife to use...which this needs to be]
I'm thinking of returning the ram and getting something that is on the QVL.
I still need 32GB. [which I guess is hard to push for the memory controller].
I don't want to run 4x 8GB sticks because I want the option to go to 64GB in the future (I plan on running this for a long time like my last system (8+years)).
I'd also like to keep running 3733/1866.
So I'm thinking QVL 3600-4000 16GB sticks are best for my application?
https://www.asrock.com/mb/AMD/X570 Taichi/#MemoryMS
Current RAM:
$190 - Newegg - G.Skill Trident Z Neo F4-3600C16D-32GTZNC
Think I have it narrowed down:
$285 - Amazon - 32GB - 3600 - C17 - HX436C17PB3K2/32
$347 - Newegg - 64GB - 3600 - C17 - F4-3600C17Q-64GTZSW
$280 - Newegg - 32GB - 3600 - C16 - CMK32GX4M2K3600C16
$425 - Newegg - 32GB - 3800 - C19 - CMK128GX4M8X3800C19
I'm leaning toward the CORSAIR Vengeance LPX.
What would you get and why?
- Joined
- Dec 19, 2012
If I were buying them I would go for something like this, Samsung B-Die typically is the easiest to work with. https://www.newegg.com/g-skill-32gb...820232903?Item=N82E16820232903&quicklink=true
OP
- Joined
- Apr 27, 2020
- Thread Starter
- #50
It's kinda why I got the Neo set. I mean it has been OCing okay... Just needs retraining every time
That's also what Buildzoid says when he was doing his FLCK/MCLK sync/unsync comparisons in the videos.
So you'd simply upgrade to Royal series even though it's not on QVL?
- Joined
- Jul 31, 2005
Maybe these 16GB NEO sticks??... Although there shouldn't be much difference between these and the B-die Royals.
https://www.newegg.com/g-skill-32gb-288-pin-ddr4-sdram/p/N82E16820232860?Item=N82E16820232860
https://www.newegg.com/g-skill-64gb-288-pin-ddr4-sdram/p/N82E16820232864?Item=N82E16820232864
https://www.newegg.com/g-skill-32gb-288-pin-ddr4-sdram/p/N82E16820232860?Item=N82E16820232860
https://www.newegg.com/g-skill-64gb-288-pin-ddr4-sdram/p/N82E16820232864?Item=N82E16820232864
- Joined
- Dec 19, 2012
I've been working with and reviewing Zen since it launched. I'm also a bencher. I have tried a few different kits and I always come back to Samsung since it just seems to work. I can put two different sets in a system and it'll boot up just fine. I realize there's a bit of a premium but to me, it's worth it. Any other kit I have had I have sold and the only thing left in the stable is Samsung B-die.
I have to ask about the RAM training that you keep bringing up. Is this just a long boot time or is it reverting to 21333 and you need to reboot?
I have to ask about the RAM training that you keep bringing up. Is this just a long boot time or is it reverting to 21333 and you need to reboot?
OP
- Joined
- Apr 27, 2020
- Thread Starter
- #53
Sometimes long boot, sometimes no post and reset to boot, sometimes revert to 2133. It's inconsistent. It's also weird that I can run memory stress tests (p95) and it is stable, but posting/booting is always difficult.
I looked at NEO vs Royal same timings and can't find a consensus. 20$ is worth it to me if they're binning Royal better so I never have to deal with this again.
- Joined
- Dec 18, 2000
I don't see any point to what you're going through with the RAM. Just run it at the 3600/1800 MCLK/FCLK XMP setting. I have cheap G.Skill/Hynix DDR4-3600 running at 16-19-19-19 with no issues with my 3900X/MSI B450M combo. There's very little if anything to be gained pushing the infinity fabric a bit higher in real world apps. Playing with secondary timings is also largely a waste of time.
- Joined
- Dec 19, 2012
Are there options in that BIOS to set the boot voltage for the RAM? I know they have it on CHVIII it really helps.
Like Dave said pushing that much for little gain isn't worth the headache and spending more $$$ on RAM. If it'll work consistently at 3600/1800 why not just go with that
OP
- Joined
- Apr 27, 2020
- Thread Starter
- #57
Because even 3600/1800 speeds don't train properly.
Not that I can find in the taichi.
I really wish I knew everything I know now about Taichi, lots of reviews said Taichi is good [which it is]. But I'm wishing I had anything else now...
See above. Just restarted and 3600/1800 1.35v stock everything rebooted into windows with 2133 ram speeds...
- Joined
- Dec 19, 2012
I feel we should concentrate on getting 3600 to boot reliably if that fails then different modules are still an option but there are a few more tricks in the book.
Can you post a pic from BIOS of all your memory related voltages/settings.
First what are your gear down and power down settings. I would set them both to off/disabled
Can you post a pic from BIOS of all your memory related voltages/settings.
First what are your gear down and power down settings. I would set them both to off/disabled
OP
- Joined
- Apr 27, 2020
- Thread Starter
- #59
Okay thanks, give me a few days.
For whatever reason Windows needs to resync my mirrored arrays even though it was properly shut down and restarted. I think during memory training win 10 kinda boots a tiny bit and it messes up the sync. I'm tired of it writing TB of data to resync instead of just comparing so I'll have to wait for it to sync and then remove the mirrors so this stops happening. Set up some sort of automatic backup to the 2nd drive(s).
---
IIRC power down is disabled, gear down is auto. I tried to disable GDM and it doesn't train at all at any of the higher RAM speeds. I'm assuming the 1T command rate isn't happy with 3600 speed so GDM is engaging to help it out.
For whatever reason Windows needs to resync my mirrored arrays even though it was properly shut down and restarted. I think during memory training win 10 kinda boots a tiny bit and it messes up the sync. I'm tired of it writing TB of data to resync instead of just comparing so I'll have to wait for it to sync and then remove the mirrors so this stops happening. Set up some sort of automatic backup to the 2nd drive(s).
---
IIRC power down is disabled, gear down is auto. I tried to disable GDM and it doesn't train at all at any of the higher RAM speeds. I'm assuming the 1T command rate isn't happy with 3600 speed so GDM is engaging to help it out.
