AMD Zen 5 Memory Performance Scaling: When Too Fast is Bad

The adoption of DDR5 memory on modern platforms has brought significant advancements in bandwidth and data transfer rates, and AMD's Ryzen 9000 AM5 platform is no exception. As DDR5 memory becomes more mainstream, the potential for performance scaling across various memory speeds offers an opportunity to understand how different configurations impact overall system behavior. In this review, we investigate the scaling of DDR5 memory on Ryzen 9 9950X, running speeds from DDR5-8000 down to the JEDEC baseline specifications of DDR5-5600 and DDR5-4800. These tests aim to highlight the performance gains and any potential trade-offs when pushing DDR5 memory beyond the standard specifications. We're also doing testing at the interesting DDR5-6000 CL28 setting as well as DDR5-6400 with UCLK 1:1.
www.techpoweru...
Intel Core i5
Core i5-12400F
Core i5-13400F
Core i5-12600K
Core i5-13600K
Core i5-14600K
AMD Ryzen 5
Ryzen 5 8500G
Ryzen 5 5600X
Ryzen 5 7600
Ryzen 5 7600X
Ryzen 5 9600X
Intel Core i7
Core i7-12700K
Core i7-13700K
Core i7-14700K
AMD Ryzen 7
Ryzen 7 5700G
Ryzen 7 5700X
Ryzen 7 7700
Ryzen 7 7700X
Ryzen 7 9700X
Ryzen 7 5800X
Ryzen 7 5800X3D
Ryzen 7 7800X3D
Intel Core i9
Core i9-12900K
Core i9-13900K
Core i9-14900K
AMD Ryzen 9
Ryzen 9 5900X
Ryzen 9 7900
Ryzen 9 7900X
Ryzen 9 7900X3D
Ryzen 9 9900X
Ryzen 9 5950X
Ryzen 9 7950X
Ryzen 9 7950X3D
Ryzen 9 9950X

[@pf100andahalf]

I like how techpowerup gets straight to the point.

[@b1lleman]

Yes indeed it resumes very well the great article on the website.

[@squisherderheld]

With the "traditional yt-media" aka HUB, GN etc. you wouldn't even be past the sponsor anouncement . And then they have to stretch the content like crazy without getting much info across. Go TechPowerUp!

[@b1lleman]

Ahhh there's a techpowerup channel. Subscribed !

[@datsweetsansabooty]

Went straight into the video and didn’t even asked me to subscribe, so I’m subbing.

[@TechPowerUp]

Maybe THAT'S why we have so few subscribers... 😁

[@spladam3845]

I wish all tech videos could be more like Tech PowerUp. Thanks folks.

[@PadPoet]

Never knew you guys had a KZbin channel, subbed!

[@aurelion3]

Never thought you guys had a channel

[@weirdodude1173]

For my 9700X I have a DDR5-6000 CL30 kit that is running at 6200 CL30 with slightly increased voltages, it seems really good. Thumbs up!

[@julfy_god]

>faster=slower literally 1984

[@PowellCat745]

1:30 6400 is very doable on Ryzen 9000 now due to IOD improvement. Recent batches of Ryzen 7000 can also run 6400 with low SOC voltages.

[@jcgongavoe337]

Ryzen 9000 didn't change IOD at all, it is the exact same chip from 7000

[@PowellCat745]

@@jcgongavoe337 as process matures silicon quality improves. That’s why I said newer Ryzen 7000s would also benefit. Learn to read bro.

[@kasimirdenhertog3516]

My 9700X will boot with 6400, but when testing with Memtest86+ it fails pretty quickly. So I’m not sure if it’s better (I doubt it even) and would encourage everyone to thoroughly test their memory when fiddling with it, because being able to boot doesn’t mean all is fine.

[@PowellCat745]

@@kasimirdenhertog3516 first of all memtest86 is a terrible stress test. Whats your kit to begin with? If your kit can’t run 6400, no amount of SOC voltage will solve it.

[@kasimirdenhertog3516]

@@PowellCat745 you may have an opinion about Memtest86+, but there's a very basic thing about tests: if it fails, it ain't good. And it's a 7600 kit, I even can run it at 8200 so 6400 should be a piece of cake.

[@TheRealDlo]

I've been using DDR5 6400 CL32 for awhile and everyone said I didn't know what I was doing... 👍

[@mttrashcan-bg1ro]

I have been with my 7800X3D and it's worked flawlessly. Currently afaik the limitations are the CPUs themselves not the motherboards, which also keep getting updates over time. So if Ryzen 9000 doesn't handle it much better, the gen after should.

[@Weroleytor]

The problem is that if you can't run it at 1:1 with your CPU then 6400-32 is slower than 6000-30, not all CPUs can run 6400 at 1:1

[@TheRealDlo]

@Weroleytor True. I can run 1:1 & when I made the purchase, that was the highest speed that would work.

[@FatheredPuma81]

Literally the video I wanted. I randomly saw there was an Agesa update that improved 7000MHz+ stability and was worried because I bought some (comparatively) slow 6000MHz memory last year for $60 because it was the sweet spot.

[@StudyWithNika]

looking forward to similar ram testing videos for upcoming intel cpus ❤

[@exscape]

It's very misleading (actually strictly speaking incorrect) to say AMD officially supports DDR5-6000. The highest officially supported (guaranteed to work) speed is DDR5-5600. 6000 almost always works (1:1), but is not supported.

[@insertnamehere4419]

Irrelevent. 6000 will work.

[@exscape]

@@insertnamehere4419 I don't see how it's irrelevant. There should be a distinction between what AMD supports and what works; explicitly saying it's "officially supported" when it's actually explicitly *not* officially supported isn't great.

[@insertnamehere4419]

@@exscape This has been how it works for a long time. Were you using 2666 memory with AM4 just because it said thats the official support limit? You're doing it wrong.

[@exscape]

@@insertnamehere4419 I never said anything of the sort. I said the video's claim that DDR5-6000 is officially supported is wrong, nothing about what does I use, nor what I recommend. And yes, I'm using faster RAM than is officially supported (and recommend DDR5-6000 CL30 for most people on AM5).

[@andersjjensen]

5600 is supported in 4 DIMM configuration. You can always run faster in a 2 DIMM configuration. And no, 6000CL30 EXPO doesn't "almost always" work. It always works. And AMD sends 6000CL30 kits with boxed processors to reviewers. They would look pretty damn stupid if that didn't work.

[@AceStrife]

The memory timings being all over the place does not make this a 1:1 test between any single instance, especially the 5600 with massively loose timings compared to 6000 (CL especially). I get it's extremely difficult to find a memory kit capable of the same timings at higher speeds, but still, this variance does impact results, quite significantly in some cases.

[@RobloxianX]

Good thing that G.Skill just released it's DDR5-6400 CL30 kits, it'll be perfect for a Ryzen 7 9800X3D and RTX 5090 build

[@AceStrife]

Perfect for parts that aren't out yet with no testing done on them, yep.

[@read6281]

This video fails to mention infinity fabric, there is a noticeable dip in latency when running FCLK at exactly 2/3 UCLK so that would be 2133 @ 3200 UCLK or 6400 MT/s (2133 is about where AMD FCLK tops out at which is why you never see anyone going for 7200-8000 etc.) Edit: I saw in a comment that you guys did your tests @2100, try bumping to 2133 @ 6400 and you should get your best results by far

[@kastrat0r]

So whats best for 3100 UCLK (6200MT)? 2066 FCLK, right?

[@read6281]

@@kastrat0r Yes 2066 would be it, one thing to keep in mind is that 2133 is a faster frequency in general so at a certain point it becomes better to to just run a higher frequency rather than the 2/3 ratio. Particularly at like 3000 UCLK (6000 MT/s) 2133 would still be faster even though 2000 FCLK has the advantage of being synced

[@kevinerbs2778]

​@@read6281Thats why a little fine tuning with 8,000mhz ram would out do the sync'd 6,000/6,200/6,400mhz Ram. You can see it on some test the 8,000 needs more bandwidth through the infinity fabric speed since its not in sync. The bandwidth need out weighs the latency hit at that point.

[@MikeBob2023]

Brevity FTW. Thank you! 👍🏼

[@berke__]

6400 mhz cl30 / 6200 mhz cl28 👌

[@Crossfire2003]

When running my Kingston memory kit at EXPO-6400MT/s CL32, or at EXPO-6000MT/s CL30 with my Ryzen 9 9950X, MemTest86 would display errors, so I had to drop the frequency at 5800MT/s while keeping the EXPO-6000 timings (CL30-36-36-80, 1.4V) in order to have a 100% stable system. I guess, there's a reason why AMD only supports, officially, DDR5-5600...

[@tanmay5570]

most detailed ddr5 scaling I have ever seen. Thank you

[@tvshacker]

Now with anandtech gone, all rests on techpowerup shoulders

[@mikeramos91]

Upgraded from 32gb 6000mhz cl36 to 64gb 6600mhz cl32 & noticed a difference.

[@MouseSpart]

my experience with ram : i had 12900k with ddr6200 cl34, upgraded to 14900k with ddr6400 cl32 and after upgrading again to 7800x3d i didnt change the ram and i had problems due to xmp. i bought ddr6000 cl30 expo and now its fine.

[@syncmonism]

Great video. Usually you don't see such accurate, useful, and empirically backed up information about such a complex topic in such a short and concise video.

[@Nintendokater]

0:06 "Fast" is a very unsharp word. Fast in terms of clock rate, or in FP32 performance? Anyway, depending on the definition, there are several models that are "faster"

[@marknikolich6986]

AES Encryption is accelerated with higher bandwidth so technically 8000 is still better 🙂 But in all honesty its not worth the money for the average user, but for a power user depending on what programs were tested, I'm not sure what techpowerup used for every single setting and test but my tests show up to 25% performance increase if latency is balanced between the optimum frequencies ie. 6400 and 8000. but some applications and even games can also have a negative hit so it depends on the owners use case. For example, a single PC streaming setup using cpu rendering, I seen a performance boost up to 10% for 1% lows in games if the ccd's are split with process lasso with the Game & OBS, since bandwidth is increased overall thorouput is also inscreased. So my conclusion is if you have the money and need the bandwidth do 8000, but if not, then you just don't waste your time. It is very hard to optimized and tighten the timings with Ryzen at 8000 to the point I deemed it not really worth it lol. Cheers all.

[@FiveMissiles]

on your guys website you have the wrong dimensions for the asus strix 1080, you guys have the pcb dimensions but not thr cooler dimensions

[@tanmayhembram9249]

Straight to the point. Thanks.

[@QuaK3RRR]

did u try fclk 2000 uclk 2000 and mclk 4000, memspeed 8000 mts? so fclk = uclk ?

[@joshua7563]

DDR4 was so clear to me with the timings and cas latency but with ddr5 I dont have that

[@ФедяКрюков-в6ь]

Same thing, basically

[@vulcan4d]

What about 6200 with tight timings.

[@CrashBashL]

Is it safe to say that the KLEVV Bolt V 32GB (2x16GB) DDR5 Gaming RAM 6400MT/s CL32 Memory Module Kit are the best ?!

[@insertnamehere4419]

No? Literally never heard of "Klevv"

[@CrashBashL]

@@insertnamehere4419 Then do some research on Google.

[@pilvisus]

If ddr5 6000 cl32 shows 2990 in 1 memory othervis 3000 then what it shows or ...?

[@PerJohansson-Xeizo]

I have a launch day 7950X from back when, I have been running it 6400c32/2133(1:1) pretty much from the start. SOC at 1.24V, VDDP 1.0V, it's super stable. Wouldn't expect less from a new CPU, but I'm waiting for the 9950X3D now with more or less confirmed dual 3D Vcache.

[@djnes2k7]

So is the ram number they referring actuall mega transfers (MT/s) so the both clocks (infinity fabric) should be the same or ram doubled? Am I getting that right?

[@TechPowerUp]

Memory controller clock and memory clock should match. Infinity fabric is different, running at ~2100 in our tests.

[@gex581990]

AMD always benefits most from low latency, has been that way since Zen 1. Nothing new. Obviously it also doesn't do as well in "Gear 2" because it's essentially raising lots of latency for a little bit more speed.

[@irataoforfun]

I read the complete article, but the video resume beatifully , thanks for the work.

[@alexmannen1991]

can u do a followup test of extreme timings can u get 6400 to 26 timing with 1.6 v and some heatsink + fan

[@Pillokun]

but, what are the other cpus running when it comes to memory settings? I got rid of all mine more expensive lga1700 cpus because the 12700k with 7200c34 at 7800c36 is just as fast, even my 7800x3d with 7800x3d is not my prime mp/fps system because of that and it relegated to race sims and youtube while I game on they 12700k.

[@TechPowerUp]

For DDR5 all systems are tested with DDR5-6000 36-36-36-76

[@rodrigorras]

how about FCLK speed at 6400 ?

[@FireCestina1200]

Are you blind ?

[@TechPowerUp]

FCLK was set to "Auto", which resulted in 2100 MHz for all these tests. This is close to the limit of the CPU, 2200 MHz is not possible

[@kevinerbs2778]

@@TechPowerUp meaning 2,000mhz be average 2,100mhz a better than average 2,133mhz lucky bin 2,2000mhz GOLDEN FIND!

[@PowellCat745]

@@kevinerbs2778most can run 2167mhz without issues.

[@PowellCat745]

@@TechPowerUpfor 6000 and 8000 you should run FCLK 2000MHz. For 6400 you should run 2133MHz to minimize latency.

[@leoariasyt]

DDR5 5600 mts is price/performance, DDR5 6000 mts is the sweet spot.

[@ФедяКрюков-в6ь]

Ddr5 5600 could be overclocked to 6000...

[@GrumpyGrndad]

And the best gaming cpu 7800x3d is left out... Is this to make intel look better? :D

[@TechPowerUp]

If you check the article you'll see the 7800x3D there. Was too high on the chart in the gaming test to show and too low in the application 😉 You can also see it in our 9950X review video. kzbin.info/www/bejne/imOxiGOCn8ucaZY

[@GrumpyGrndad]

@@TechPowerUp thx

[@gyldesign8388]

take a 4090 for 720p gaming its like take a formula one for test into bottleneck trafic.

[@PolskiJaszczomb]

Wow, somebody discovered, that IF clock is the bottleneck, who'd thought about that, ever since Zen 3? Surely nobody.

[@TechPowerUp]

This is a separate clock than the fabric speed.

[@PolskiJaszczomb]

@@TechPowerUp a verification of uncoupled vs coupled affecting performance would be appropriate.

[@hytalefanboi7471]

Maybe someday a manufacturer manages to create DDR5 12000MT memory lol

[@PorinaAdventures]

I like this video format! Keep it up. The written site is already a go to for me.

[@fffhunter7765]

Just what i needed 10x sub +1

[@Rentta]

Would this be the case with 7000 series cpu's too ?

[@TheRealEtaoinShrdlu]

Every generation has its own optimal memory speed. 5600 MT/s is optimal for Zen4 if I remember correctly.

[@26Guenter]

DDR5 6000 is best for 7000. Although some comments are saying they can run 6400 so maybe if you play around you can run it.

[@themusesquad8554]

In gaming their is also single and doul rank differences as well as with timing.

[@ZERARCHIVE2023]

Grabbed 6000mhz for better compatibility, don't really care about RAM speed, but quantity ? Yes.

[@pankothompson5903]

excellent well done

[@BenchmarkingTom]

Hello, what do the voltages for the 6400MHz and 8000MHz profiles look like? For example, if the 8000 (1:2) required significantly lower voltage than the 6400 (1:1), wouldn't it be more efficient, especially if the performance difference between them is only 0.1%?

[@TechPowerUp]

www.techpowerup.com/review/ddr5-memory-performance-scaling-with-amd-zen-5/20.html There isn't a lot of difference between any of them, though 6400 vs 8000 they are pretty much tied.

[@Matlockization]

I don't like the background music when someone is talking to us. What's the native for DDR4 ?

[@justinfoster1040]

Great video, not wasting time, clear and effective. No sensationalism just info, thanks!

[@ahmed_x5597]

I don't get it why this channel doesn't get many views/subs their content is good and straight to the point

[@toddsimone7182]

My favorite benchmark charts and website for all tech reviews now has a youtube channel? Sign me up!

[@andersjjensen]

I'm sorry, but 1 percentage point increase is not "decent". Unless you know that CPU you buy will do 6400 in 1:1 without a lot of voltage there is no point in shelling out the extra cash for the memory.

[@ThaexakaMavro]

I run 64gb of 5600Mt/s cl34 pretty happy with the middle ground memory size / speed

[@AsthmaQueen]

great work and coverage

[@noobnook2890]

All we need is DDR5 12000 lol

[@mikeramos91]

At that point it will be called DDR6 🤔

[@benuten3907]

8k mt's only work best with Intel..

[@raminpro9765]

6400 is sweet spot on 9000cpu(many youtubers use 6000mhz)

[@ThaexakaMavro]

barely any gains vs 6000mhz so no the sweet spot is still 6000

[@ReallySadProfessor]

@@ThaexakaMavro Can you overclock 5600Mhz kits to 6000? Or does your kit need to be made specifically to run that, just asking. Because I'm upgrading to a 9950x soon, and got a CL46 5600mhz kit.

[@ivantsipr]

​@@ReallySadProfessoryou have to manually overclock it and stress test it , also which memory chip the kit has is the determining factor, Hynix memory is the best available and has the highest chance of clocking to 6000 (since the default spec of your kit is 5600 means is a low binned hynix or is another brand), if it is Samsung or Crucial memory they max 6000, kind of, but with horrible timings, and most of the DDR5 performance comes out of the timings, that's why for high speed kits or overclocking you want nothing else than Hynix.

[@ThaexakaMavro]

@@ReallySadProfessor yes most 5600mhz kit will overclock to 6000mhz quite easily .

[@ReallySadProfessor]

@@ThaexakaMavro Ok I got two completely diffy answers and contexts....I do believe Crucial makes quality RAM so even with bad timings, I should be able to push some voltage into it and force 6000. It runs 1.1V native anyways. thanks for the answer

[@walter1824]

Is it possible to "futureproof" RAM for upcoming CPU's

[@26Guenter]

No, you don't have a guarantee on the speeds supported or could run higher speeds than what you bought.

[@Ray88G]

But this is only gaming . Productivity workloads will benefit from greater memory bandwidth

[@TechPowerUp]

Productivity is the last part of the video.

[@carlchristenisnes6763]

informative - perfect , thanks

[@Games_and_Tech]

I have my 7800x3d ans my 64 gb memory kit set yo 6400 MT/S cl32 on 1:1 ratio 😊

[@airzeroHD]

LoL my ryzen 9 9950x runs perfect CL32 6400

[@ThaexakaMavro]

CooL

[@insertnamehere4419]

K? That doesnt refute the point in the video.

[@xblur17]

So does theirs, what's your point kiddo

[@MrAckers75]

Make sure you get Hynix a die ….l

[@inqizzo]

Its best to buy very fast ram, for example lexar ares 7200cl34 and clock it down to 6000cl28. And then because its a good bin, most of them can go up to 6000cl26 if you bump the voltage to 1.4 from 1.35

[@GrainGrown]

​@@inqizzo *It's

[@ivantsipr]

​@@DesoloVirhigh capacity and high speed have a inverse correlation, for high speeds you want 2 sticks only, 4 sticks is too much stress for the memory controller, for high capacity (>96gb) you need 4 sticks, so you have to choose high capacity or high speed, you can't have both.

[@ivantsipr]

@@DesoloVir 2 x 48gb = 96gb, read my comment

[@Reza1984_]

This could be a half an hour video on other tech channels lol

[@TheSkepticSkwerl]

What was the point of this video? Seriously. You explained timings. But didn't show results. Or anything. This was a useless video.

[@FireCestina1200]

Just get a 6GT/s CL28-30-30-66-104 if you can.