On the single rank vs double rank test the 4x8 GB DIMM's will be dual-rank as far as the memory controller is concerned because you have two single rank DIMM's per controller. The only true single rank is the 2x8 test because its a single rank DIMM per memory controller. This is why the single rank(2x8GB) is the only one that suffers while the other three are within a 2% margin of error. The short version is to populate all 4 DIMM's on Ryzen because it seems to offer better performance regardless if you have dual or quad rank per memory controller. This has been the case since first generation Ryzen as well.
Maybe a review from 1usmus before publishing would have been helpful 🙄
The Aorus X470 Gaming 7 WIFI is a board with T-Topology layout...
These recommendations are not going to be good for boards with Daisy chain layout (almost all X570).
Especially the 4 sticks with Dual Rank, hard to have them working with low timings and/or high clock.
Also it should have been mentioned Gear Down mode, what it means, what it does, etc
Maybe a review from 1usmus before publishing would have been helpful 🙄
The Aorus X470 Gaming 7 WIFI is a board with T-Topology layout...
These recommendations are not going to be good for boards with Daisy chain layout (almost all X570).
Especially the 4 sticks with Dual Rank, hard to have them working with low timings and/or high clock.
Also it should have been mentioned Gear Down mode, what it means, what it does, etc
Very good point about the change over to daily chain.
Interesting... seems Zen2 is a lot less bandwidth-hungry than 1st gen.
Also, great to see the mins and average FPS. I hope to see that more in future benchmarks.
The point here is that i can keep my 4x8 cl14 if i swap from 2700X to another ryzen on the x470 board, right? seems that i gain nothing optimizing memories
Great article thank you,
One of the most in depth I've seen and puts the memory argument to rest. Perhaps the only thing I would've like to see would've been minimum frames on 1440p but I assume the 1080p highlights that benefit that would partially translate to 1440p.
Much appreciated!
Extremely interesting & I have to say excellent article, very in-depth credit to Krzysztof Hukalowicz for the time it woould have taken to gether all these results!
The point here is that i can keep my 4x8 cl14 if i swap from 2700X to another ryzen on the x470 board, right? seems that i gain nothing optimizing memories
Kick ass write up! To see that today how things have changed where more does actually equal better?!?
I truly remember we're having more dimms populated could potentially impact gaming performance in a negative way. You had to have some fast ram to truly make the difference then.
@Krzysztof Hukalowicz great job putting this together and @Hilbert Hagedoorn of course.
This is quite a bit of information thank you so much guys.
These recommendations are not going to be good for boards with Daisy chain layout (almost all X570).
Especially the 4 sticks with Dual Rank, hard to have them working with low timings and/or high clock.
Daisy chain board will pose problems with 4 sticks only at really high frequency (higher than the IF can go).
I have an aorus master (Daisy chain) with 4 stick of DR 16gb Samsung B die (64 total, need it for photo editing). it run perfectly at 3733 (1 to 1 if ratio) at Cl 16-16-16-16 (1.37v). And yes 8gb SR stick could probably go to cl 15 (maybe 14) but that's not because of the board layout but because of the stick themselves.
Daisy chain board will pose problems with 4 sticks only at really high frequency (higher than the IF can go).
I have an aorus master (Daisy chain) with 4 stick of DR 16gb Samsung B die (64 total, need it for photo editing). it run perfectly at 3733 (1 to 1 if ratio) at Cl 16-16-16-16 (1.37v). And yes 8gb SR stick could probably go to cl 15 (maybe 14) but that's not because of the board layout but because of the stick themselves.
The different board layout changes the termination; T-Topology is 50/50 while Daisy chain is about 75/25.
Doesn't change how low you can go with timings but how reliable they can work under borderline conditions, when you want to get the best out of them.
Running Samsung B-die at 3733 CL16 is super easy; but if you want to tighter the timings or raise the clock you'll reach the limit much before than with a T-Topology.
Good B-die can go down to 3800 CL12 but not DR, not 4 sticks in a Daisy chain.
If you have much cheaper RAM, like my G.Skill modules using Hynix DJR, stability problems will come out much earlier with lower speed and higher timings.
It's a good article but I'd like to see a comparison with a similar X470 with Daisy chain topology and also GDM enabled/disabled.
On my AORUS Master these G.Skill modules are faster in CL16 1T GDM-Enabled than CL15 1T GDM-Disabled.
Maybe a review from 1usmus before publishing would have been helpful 🙄
The Aorus X470 Gaming 7 WIFI is a board with T-Topology layout...
These recommendations are not going to be good for boards with Daisy chain layout (almost all X570).
Especially the 4 sticks with Dual Rank, hard to have them working with low timings and/or high clock.
Also it should have been mentioned Gear Down mode, what it means, what it does, etc
For most of the users it won't be a problem, as for the usual 3600 MHz, even the Daisy Chain MoBo will do good work with 4 sticks. It's not an extreme overclockers guide, but rather something for "Average Joe" type of guys.
schmidtbag:
Interesting... seems Zen2 is a lot less bandwidth-hungry than 1st gen.
Also, great to see the mins and average FPS. I hope to see that more in future benchmarks.
I'll try to follow that routine in the next reviews as well.
Nictron:
Great article thank you,
One of the most in depth I've seen and puts the memory argument to rest. Perhaps the only thing I would've like to see would've been minimum frames on 1440p but I assume the 1080p highlights that benefit that would partially translate to 1440p.
Much appreciated!
Thanks. 1440p would make it flatter, as usually, the GPU would be the bottleneck (even RTX 2080 Ti) in some of the games/scenarios.
chinobino:
Extremely interesting & I have to say excellent article, very in-depth credit to Krzysztof Hukalowicz for the time it woould have taken to gether all these results!
Very nice article indeed HH:) and very relevant to me as i'm in the process of upgrading my PC and ram has been the grey area for me up untill today, so Thanks again 🙂
This test is one year too late.. but better late then never, huh?
I've been touting on various forums (including this one) about optimized RAM settings on Zen 2 since launch last year, but usually hit a wall of fanboyism and naysayers and gave up.
DDR4-3800 CL16 1:1 can be achieved on some chips and not on others... Infinity Fabric silicon "lottery".
I am lucky that mine IS stable at IF = 1900 Mhz, and loving it with Crucial Ballistix Elite memory tuned to the edge.
I also have 4 sticks, which enables Dual-Rank access, improving the performance that tiny bit further.
The effect of fast and tuned RAM is VERY noticeable in some applications (especially certain games), but also in various professional applications that use a lot of RAM in a chaotic fashion.
The opposite type of software that has no variance is Cinebench, which seems to fit inside the CPU cache resulting in no performance gains at all from faster RAM.
---------
@Hilbert Hagedoorn
It would be interesting to add another chart to compare, for example, a 3800X + PBO & scalar X7, with tuned RAM vs 10700K with 5Ghz OC (which probably anyone can achieve).
I'm quite sure the "gap" between the two is very tiny when not intentionally crippling AMD chip with default loose memory timings and low clocks.
As I was for years using mainly the Intel, AMD Ryzen platform was a new experience for me (at least in making experiments in the RAM department) - so I've set it according to the SAFE recommendations from Ryzen DRAM Calculator for B-Die memories. So nothing spectacular here, as I would have to have spent much more time to dig into the super-hiper-best settings there.
feboi:
`d really appreciate if someone can experiment and make a summary.
Already this article took a couple of nights from my life (and I have a normal job + wife + 2 kids), please - have mercy on me, haha.
squalles:
weird, this is the unique article in the world who shows 3733mhz not causing major improvements
2% on average and 3.3% improvement is too low? Please show me the article where it's significantly better. How much should it be - 10%? (when you divide 3733/3600 you get 3.7% more) . Of course I'm comparing the 3733 CL16 vs 3600 CL16.
I also have 4 sticks, which enables Dual-Rank access, improving the performance that tiny bit further. The effect of fast and tuned RAM is VERY noticeable in some applications (especially certain games), but also in various professional applications that use a lot of RAM in a chaotic fashion. The opposite type of software that has no variance is Cinebench, which seems to fit inside the CPU cache resulting in no performance gains at all from faster RAM. --------- @Hilbert Hagedoorn It would be interesting to add another chart to compare, for example, a 3800X + PBO & scalar X7, with tuned RAM vs 10700K with 5Ghz OC (which probably anyone can achieve). I'm quite sure the "gap" between the two is very tiny when not intentionally crippling AMD chip with default loose memory timings and low clocks.