AMD Ryzen Memory Tweaking & Overclocking Guide

[Deleted member 50521]

Amazing guide for RyZen owners? I assume it applies to Threadripper users' as well?

Also I would LOVE you write an Intel version as well.

 

[Vycyous]

@1usmus Thank you for taking the time to write this article (and develop the DRAM Calculator for Ryzen). I plan to read through it more thoroughly this weekend.

I didn't see this mentioned, and I haven't seen anyone else mention this, but the biggest problem I've found is knowing which memory kits to buy (not just for Ryzen, but also for Intel). It's almost a complete crapshoot. There are only a handful of memory kits that I know should be, for example, all-but-guaranteed single-rank Samsung B-die. Through deductive reasoning, reading datasheets and personal experience, I've determined those kits to be 8 GB DIMMs with speeds/timings of 2400 MT/s C10, 3200 MT/s C14, 3600 MT/s C15, 4133 MT/s C17, and mostly anything above 4133 MT/s, whether it's C18 or C19. However, that leaves a lot of memory kits in the unknown category.

I was surprised to see the 3600 MT/s C19 Hynix kit that was mentioned because I never would have thought about purchasing one of the highest CAS latency 3600 MT/s kits. For the most part, I've had good luck with Hynix for Ryzen (though not as good as single-rank Samsung B-die, of course), so I'm not afraid to purchase another kit with those ICs. I do avoid Micron, though.

I haven't had very good luck with Micron (as far as overclocking goes) on either AMD or Intel platforms. Maybe that will change with the new Micron H/E-die (16 nm) ICs mentioned in the article. Fortunately, Micron is a little easier to avoid; don't buy Crucial/Ballistix-branded memory. Of course, nearly any kit below roughly 3200 MT/s C16 could be using Micron ICs (and a lot of them are). Could this memory kit possibly be using some of the new H/E-die (16 nm) ICs? At the price they're currently asking, I'd rather buy a kit of all-but-guaranteed Samsung B-die for the same or less.

 

[Zubasa]

There are actually some Crucial kits that uses Samsung B-Die, but so far from what I heard they tend to be pretty disapointing B-Die.
I guess it is no a surprise, given how loose the timings are out of the box.

 

[Vycyous]

There are actually some Crucial kits that uses Samsung B-Die, but so far from what I heard they tend to be pretty dissapointing B-Die.
I guess it is no a surprise, given how loose the timings are out of the box.

That's interesting, and I guess that sort of confirms that Crucial/Micron really didn't have anything that could clock up beyond roughly 3200 MT/s. I guess that might be changing with their new 16 nm ICs. Those timings really aren't that bad for 3466 MT/s, although it almost certainly means it's a slightly lower-binned B-die chip (just as you said, although there is definitely a lot worse B-die). I know Corsair, and probably others, also use Samsung B-die in some of their 3466 MT/s C16 kits.

 

[Zubasa]

That's interesting, and I guess that sort of confirms that Crucial/Micron really didn't have anything that could clock up beyond roughly 3200 MT/s. I guess that might be changing with their new 16 nm ICs. Those timings really aren't that bad for 3466 MT/s, although it almost certainly means it's a slightly lower-binned B-die chip (just as you said, although there is definitely a lot worse B-die). I know Corsair, and probably others, also use Samsung B-die in some of their 3466 MT/s C16 kits.

The 3600 version you linked runs at similar timings, so those should be strait up better than the 3466 version.
3466 CL16 is not horrible, but I rather grab some 3200 CL14 which has better compatiblilty and better latency.

 

[Vayra86]

Awesome guide, great detail and info. Wow.

 

[Countryside]

Beautiful guide and well done, keep up the good work.

 

[TheEmptyCrazyHead]

Been reading for a while and finally did a registration to say applauds and huge thanks for this article! @1usmus I bow before your patience and punctuality!

 

[mtcn77]

This article is almost up to date with hardware engineering pov. Great stuff.

 

[_larry]

Very good write up!!
I have a set of Corsair 2x8GB 3200mhz and an Asus Prime 350b-PLUS mobo, but the RAM kit is not listed in the vendor certified list...
Therefore, I have only been able to get them to 3066mhz stable..
I run them at 3000mhz for 24/7 stability. Hopefully I can use this guide to get them into the 3200mhz+ range.
How do you find out what type of memory your RAM has? (Samsung, Elpida, Hynx, etc?)
They are Corsair Vengence LPX
CMK16GX4M2B3200C16

 

[biffzinker]

How do you find out what type of memory your RAM has? (Samsung, Elpida, Hynx, etc?)

Try using Thaiphoon Burner freeware version - http://www.softnology.biz/

Have it read one of the SPD ROMs off your two sticks then click the report button. Take a screenshot by clicking on the bottom screenshot then post it here if you want.

 

[1usmus]

@1usmus Can you explain what Multi Rank memory is? I have always assumed Single Rank and Dual Rank were synonymous with single sided and dual sided, but I guess this isn't the case.

MR ( Multi Rank) = 4 Single Rank DIMM's

Nice article!

Uncached startup times of bigger applications like games / photoshop or comparable tools would be interesting to see the differences as well!

( start from SSD ofc, HDD would negate any considerable effect that is significant enough to be compared )

In Photoshop, there is almost no increase performance. Complete analogy of the results in render packages.
There will be more detailed research and I will publish it with the debut of new processors. There will be a detailed comparison of all that may be of interest to the public.

Re-read my post, "... two dimms ... two dimms ...". Btw however we found no difference between 2x8 and 4x8 and 2x16 on X470 with these particular dimms, how's that for curious?! 4x16 is another story though.

Unfortunately I do not have kits to show you results, but the frequency of 3200 is a reality for 4*16

Absolutely great article @1usmus , thank you and tpu as such guide and information is well received and needed by the community , appreciate it your hard work.

@1usmus , i need your help my friend , i want to upgrade to a 32GB 2x16GB memory kit , I'm currently running right now a G.Skill FlareX 3200Mhz cas14 kit 16GB 2x8GB @ 3533 Cas14 , but i do not want to loose too much performance on my Ryzen 2700x and Asus CH7 as i use this pc 24/7 for everything and is my Gaming warrior / Encoding , decoding large files / large video and photos processing , etc ...

I was looking at some kits at newegg and the price range wildy and sometimes i cannot figure it out wish memory ICs are used on this kits since no information can be found. Can you recommend me a good 32GB 16x2GB kit without breaking the bank and squeeze great performance out of them ? Thank you in advanced and keep up the Awesome work !

I was looking at this kits:

https://www.newegg.com/Product/Product.aspx?Item=N82E16820232214&ignorebbr=1 - Sammys B-dies / $279US

https://www.newegg.com/Product/Product.aspx?Item=N82E16820232660 - Hynix MFR - $169US

https://www.newegg.com/Product/Product.aspx?Item=N82E16820225032 - Micron ??? - $194US

i use kit G.Skill F4-3000c14D-32GTZR (2*16) , great memory

https://www.newegg.com/Product/Product.aspx?Item=N82E16820232214&ignorebbr=1 - Sammys B-dies / $279US , it is similar to mine

Ah, I thought this memory dependency for Ryzen was fixed with updates.
Don't have time to read the article right now, but can tell it is very informative.

Do you think memory speeds will be just as important for the next gen AMD CPUs (Matisse 3000 series)? I believe they still use Infinity Fabric

I prepared inside information about new processors and their features, the news will be published in the coming days

A great article for sure, very detailed, and I love that. But at the end of the day, at least in games in real world scenarios, you won't notice the difference sadly. It really doesn't make a huge difference that people say it does, and the reality is high clock memory is not super important. A good 2933, or even 2666 kit will be good enough.

there are streamers and there are owners of monitors 144hz, also tuning directly affects the minimum FPS

Amazing guide for RyZen owners? I assume it applies to Threadripper users' as well?

Also I would LOVE you write an Intel version as well.

Yes, this guide is fully compatible with platform X399
For Intel, this is possible, I will think about creating a similar manual

Guys, I'm glad to hear what you like, at the moment I am working on this guide, I am writing new pages, and personal answers will be a little later

 

[Silent_Scone]

MR ( Multi Rank) = 4 Single Rank DIMM's



I prepared inside information about new processors and their features, the news will be published in the coming days



there are streamers and there are owners of monitors 144hz, also tuning directly affects the minimum FPS



Yes, this guide is fully compatible with platform X399
For Intel, this is possible, I will think about creating a similar manual

Guys, I'm glad to hear what you like, at the moment I am working on this guide, I am writing new pages, and personal answers will be a little later

Great, while you’re at it please add why PRECHARGE needs to wait tCAS-1, when tras min can be only tRCD+(bl-1), otherwise you risk confusing people trying to understand relationships. Your article is way too short there. You can also explain why tWR should be considered within a read tRC cycle, because that needs clarification, too. Looking forward to additional info!

Also, please explain why most white papers include tCAS in tras min, and it’s outside the tBL window? And if tCAS does not matter here, where and how does it actually affect performance? These are very important questions your guide must answer if you are touching on these subjects and changing what has been explained by most DRAM engineers.

Please see my other posts for more context.

 

[1usmus]

Great, while you’re at it please add why PRECHARGE needs to wait tCAS-1, when tras min can be only tRCD+(bl-1), otherwise you risk confusing people trying to understand relationships. Your article is way too short there. You can also explain why tWR should be considered within a read tRC cycle, because that needs clarification, too. Looking forward to additional info! Please see my other posts for more context.

On the page that interests you, the theory is published and the formula is published, in which one of the teams gets a negative value. This is not a mistake, only due to a negative value I can demonstrate a “savings”, since 2 commands are executed in parallel.
Presets - is the parallel execution of certain commands and the lack of margin, which can give the user up to 14% of FPS.
Looking at this preset, you should have a lot of questions, because it violates all the formulas and at the same time is absolutely stable.

 

[Silent_Scone]

Presets - is the parallel execution of certain commands and the lack of margin, which can give the user up to 14% of FPS.
Looking at this preset, you should have a lot of questions, because it violates all the formulas and at the same time is absolutely stable.

Yes, but this does not explain why PRECHARGE has to be sent at a minimum of CL-1 in this scenario, why not simply tBL-1, which is always 7. If this is not allowed here, why is it allowed in tRAS min?

1) Why do most white papers include tCAS in tRAS min, and it’s outside the tBL time window? Please see the IBM white paper I posted earlier, and also many others online show this.

2) How does tWR become active during a read RAS cycle? You’re using it as part of your calculation, please explain why and how it is important when there are no write commands sent.

These situations must be explained clearly in your guide, otherwise, they don’t quite make sense.

 

[Silent_Scone]

On the page that interests you, the theory is published and the formula is published, in which one of the teams gets a negative value. This is not a mistake, only due to a negative value I can demonstrate a “savings”, since 2 commands are executed in parallel.
Presets - is the parallel execution of certain commands and the lack of margin, which can give the user up to 14% of FPS.
Looking at this preset, you should have a lot of questions, because it violates all the formulas and at the same time is absolutely stable.

Ok, it still doesn’t explain why precharge cannot be tBL-1 in both scenarios

Your IP address is identical to the address from which a person with a fragile childish psyche has poisoned a lot of dirt . I will not enter into dialogue with such people.

I knew you’d dodge me if I asked directly, so yes, I created an alternative account. But, in no way is any of this childish.

The questions are valid, 1usmus. These are valid ones that bring the formula into question. They go against almost every technical document available online. As the author of this paper, I think it's your responsibility to clarify these things. This isn't something that should offend, so if it has, I apologise (I've since asked for the accounts to be merged). These topics aren't easy to cover, so the theory needs to be solid.

 

[_larry]

Try using Thaiphoon Burner freeware version - http://www.softnology.biz/

Have it read one of the SPD ROMs off your two sticks then click the report button. Take a screenshot by clicking on the bottom screenshot then post it here if you want.

Looks like they are Samsung! Let's get these puppies up in clock speed, shall we?

Should I just crank up the voltage to get to 3200mhz? What other parameters are best for stability?

 

[biffzinker]

View attachment 119244

Looks like they are Samsung! Let's get these puppies up in clock speed, shall we?

Should I just crank up the voltage to get to 3200mhz? What other parameters are best for stability?

But there Samsung E-die instead of the B-die that have the better overclocking potential.

 

[_larry]

But there Samsung E-die instead of the B-die that have the better overclocking potential.

Wonderful...Well, I guess I will continue to use them at a lower frequency and sell them when Zen 2 hits..

 

[Super XP]

Ah, I thought this memory dependency for Ryzen was fixed with updates.
Don't have time to read the article right now, but can tell it is very informative.

Do you think memory speeds will be just as important for the next gen AMD CPUs (Matisse 3000 series)? I believe they still use Infinity Fabric

Last I heard AMD is taking a different approach with ZEN2. There's going to be some core design changes in ZEN 2 versus the original ZEN. The changes will include a reduction in Infinity Fabric latency. Something about keeping Infinity Fabric's speed at a constant high speed. And basically doubling up everything. Also there will be 7nm chiplets and IF and a separate single 14nm to house the PCIe, IMCs, IF and so on.

 

[Alpi]

Bdie has more potencial in oc than edie. The timings are much more flexible but it's better for clocks even. However it won't cause any limit for Your Ryzen I guess. I tryed how my 2x4 edie goes in Am4 and it was able to boot and go os easily at 3600mhz. I want to test them again a bit longer soon.

 

[robot zombie]

I honestly have nothing to add here, but as someone who has been highly interested in Ryzen RAM overclocking for quite a while, this article is a huge help. It clarified so many things for me and set me straight on a thing or two - in the past I've been wrong about certain things... now when people ask I can give them the correct explanation and hopefully help them out as well.

A few things, I wish were explained in more detail, but even those parts make for great starting points into my own learning and research. I now have a lot more to think about and look into thanks to this write-up.

Seriously awesome stuff. I appreciate the effort put in by everyone involved!

 

[Zubasa]

Wonderful...Well, I guess I will continue to use them at a lower frequency and sell them when Zen 2 hits..

TBH Samsung E-Die is not that bad for daily use, you should be able to get it running @3200 using the Calculator.
Because even when running @3200, the Calculator gives you more optimized values compare to the XMP settings anyway.
You could still see some gains, and 3200 shouldn't be hard to stablize with the right settings.

 

[ThatNerdGUI]

Hello everyone, I'm new around here.

I currently have 4x Corsair Vengeance RGB Pro on a STRIX X470-F mobo and due to having issues trying to get them to run at the XMP rated speeds, I decided to give this guide a shot to see if I could at least get them to run at their base speed. I downloaded all the tools and the first thing I noticed was that Thaiphoon Burner was showing a CRC error in 3/4 of the sticks. Manual timings were working, but then default/manual settings stopped passing the RAM post when all RAM slots are populated and only pass with XMP on in bios . I exported the TB reports but I don't know what to make about them. Should I be concerned about the CRC errors?

I've uploaded the exports here in case someone could help.
Thanks.

 

[WuY276]

I have R5 2600, MSI B450 GAMING PLUS and 2x8GB Hynix AFR. It stable on 3200MHz with 16-17-18-18-34 timings, but don't work on 3466 with any settings. I've tried all the options that Ryzen DRAM calculator offers, but always get an error in Prime95. What can help?