DDR5 Memory Performance Scaling with AMD Zen 5

[Dr. Dro]

You can shrug off that idle power draw on the desktop, but I hope they can do something about it on their mobile parts.

Mobile parts shouldn't be affected (exception of some HX which use the IOD), it's the IOD's static power consumption that causes the high idle draw.

 

[kiakk]

And then at full speed, you need a nuclear power plant to feed that intel chip.

Gaming pwr. consumption is very similar between 10600K and 5600X. I benchmarked them in PUBG, both of them is ~80W avg. in PUBG with CPU benchmarking setting (720p and 70%res) with unlimited framerate.
With the gaming settings 1440p@163fps limit, 5600X/10600K consuming very simmilar, 60-65W avg. The differences is negligible.

Only Intel high end CPUs workload power is insane. But for maisntream 6c/12t gaming is very comparable between AMD and Intel. I also watched countless benchmarks and tested myself (that is why i have now 2PC)
While AMD have advance in efficiency high core count workloads, Intel have advance in idle and light load and also games which do not need many threads. I use my pc for light loads (like browsing news, youtube, forum, etc...) more than gaming, so idle and light load is important for me as well.
That is why I wrote that AMD should take care of IO die energy management.
While AMD take care of energy management in the monolitic design they should as on desktop architecure as well.

Intel power consumption is very over reacted on the internet in my experience.

 

[BoggledBeagle]

That KF560C32RSK2-96 is a nice memory kit.

I got the same kit a while ago, for some reason it is much cheaper than the 6400 MHz version - at least where I live.

Is it usable with AMD CPUs, I believe it does not have EXPO profiles in it.

 

[EmmanuelMar]

El rendimiento final entre estas Ddr 5 es escaso no igual los precios de venta de estas DDR.
Se ven influenciados por el cas a menor cas menor latencia tal como se ve en las pruebas y de hecho las 6000 mhz y las 6400 se llevan el premio....las ddr 4800 y 5600 rendirian mas de ser cas 30.
El cliente que posee una cpu amd X3D poco le afecta en juegos que usan X3D no así una cpu Intel que será más dependiente de la DDR.
Personalmente solo compraria una DDR 6000 y cas 30 si poseo un chip serie 9000 Amd pues son chips mas pensados para aprovechar esas frecuencias.

 

[ir_cow]

I got the same kit a while ago, for some reason it is much cheaper than the 6400 MHz version - at least where I live.

Is it usable with AMD CPUs, I believe it does not have EXPO profiles in it.

XMP/EXPO is interchangeable really. The only difference is EXPO has some extra values tCCD, tRRD, tWTR, tRTP, tFAW and tCCDW if the vendor chooses to include them in the EXPO profile. Otherwise the motherboard just picks i t instead like XMP. In theory EXPO would be better for both AMD and Intel, but usually those extra values are the exactly as the Motherboard would set without it anyways.

 

[HTC]

Stupid question @W1zzard : for gaming benchmarks, shouldn't you use a GPU that is less powerful in order for the system RAM to be more heavily utilized, thus better showcasing how the different timings affect the performance?

Pick an 8GB GPU and use settings that FORCE the games to resort to using the system RAM to a higher degree, by selecting games with settings that require MORE than 8GB VRAM.

If my reasoning is correct, the differences between the various timings will be exacerbated this way.


Or is my reasoning flawed?

 

[ir_cow]

@HTC just the opposite. Less powerful video card means your more likely to become GPU bound and all memory is essentially the same. Lower the graphical settings on a mid-range GPU and it has the same outcome as a higher end GPU using max settings.

 

[HTC]

@HTC just the opposite. Less powerful video card means your more likely to become GPU bound and all memory is essentially the same. Lower the graphical settings on a mid-range GPU and it has the same outcome as a higher end GPU using max settings.

But when the game is "asking" for more VRAM than the GPU has, it uses system RAM more, which is the point i'm trying to make: have games run in such situations PRECISELY BECAUSE these require more system RAM to be used than with a top GPU, hence the 8GB GPU suggestion.

THEN see how the various games performance is affected, with the different RAM timings.

 

[AnarchoPrimitiv]

Not that it's news, but Zen 5 truly retains the utterly dismal memory performance scaling past 6000 seen in Zen 4. I expect Zen 6 to be much of the same if it remains on socket AM5. Raptor Lake does see real gains when pushing memory into the 7-8K ranges.

Yes...performance for those willing to pay egrigious prices for those memory kits

 

[ir_cow]

But when the game is "asking" for more VRAM than the GPU has, it uses system RAM more, which is the point i'm trying to make: have games run in such situations PRECISELY BECAUSE these require more system RAM to be used than with a top GPU, hence the 8GB GPU suggestion.

THEN see how the various games performance is affected, with the different RAM timings.

Interesting concept. Not sure what the impact is when you go over VRAM buffer. either its a really low end card or you aren't willing to lower the settings the only scenario I can think of.

 

[doc7000]

I wonder why they used Windows 11 22H2 and not 24H2 or patched 23H2?

 

[HTC]

Interesting concept. Not sure what the impact is when you go over VRAM buffer. either its a really low end card or you aren't willing to lower the settings the only scenario I can think of.

I know from some videos that games with higher VRAM usage than the GPU has will use NOTICEABLY MORE system RAM, which is the whole point.

THEN, the comparison with the different timings can be made.

 

[CheapMeat]

I think your right HTC, that seems logical to me.

I don't think so, depends on your definition of significant though, but I'll definitely explore it, once all the pieces of the puzzle are together: X870E, Agesa, 24H2, X3D

That's great, looking forward to it W1zzard. I guess I meant to say "meaningful difference". So curious, if it's true, how the two caches on each CCD handle the RAM you tested.

 

[W1zzard]

I wonder why they used Windows 11 22H2 and not 24H2 or patched 23H2?

ah that's a typo, I used 23H2 of course. Nice find, fixed in all affected reviews (everything after 14900KS). 24H2 isn't out yet, only a beta, and Microsoft's "RTM" bs isn't worth anything these days when so many things are still broken and you get major patches on day one.

The OS was fully updated in April 2024, and then locked down, so it stays constant for the lifetime of the platform, to keep things identical for all the ~45 CPUs tested

Once X3D, Arrow Lake, 24H2 are released I will spend a few weeks full time, retesting everything, also new app versions, new BIOS, new drivers, new games

 

[_roman_]

DDR5 4800 transfers data at 4800MHz, it's not a marketing lie. Saying it's a 4800MHz clockspeed is a lie but what is a clockspeed these days anyway? There are multiple different clock domains in most modern processors with Zen using de-synced core, fabric, and memory clocks for years now. Anyone still thinking of CPUs and GPUs, even mobile phone SoCs as having a single clockspeed is stuck in the past.

? Are you serious ? go there with a scope and measure, than come back to myself and show me the nanoseconds it takes for a full cycle please. Than calculate.
Hint: frequency = 1 / time
Hint: Double Data RATE

 

[JWNoctis]

? Are you serious ? go there with a scope and measure, than come back to myself and show me the nanoseconds it takes for a full cycle please. Than calculate.
Hint: frequency = 1 / time
Hint: Double Data RATE

Unless I'm missing something significant, that statement is still correct. Double data rate meant transfer on both edges of the clock signal half the frequency, thus data transfer occurs at twice the frequency of the 2400MHz clock. AFAIK timings are also in transfers instead of clocks. EDIT: Actually no, it's in clocks. My mistake.

I'd love to have a scope that goes up to 2.4GHz and more, though.

 

[Butanding1987]

Would have wanted to see 6400 CL 28 and 8000 CL34. And why not 8200 or 8400 since it's a 9950X?

 

[jxdking]

Would have wanted to see 6400 CL 28 and 8000 CL34. And why not 8200 or 8400 since it's a 9950X?

I doubt 8000 is absolutely stable at the Hero board.

 

[ir_cow]

I doubt 8000 is absolutely stable at the Hero board.

Its possible with a nice IMC. Though I'm using the ASUS Gene because its not fully stable on any 4-slot motherboard with the CPU I have without crazy voltages I don't like using. We will soon have X870 and higher mem support.

 

[Bagerklestyne]

If I was looking at faster ram than what I have now (CL30/6000) I'd be looking for ram I could tighten @ 6000 as much as possible

 

[529th]

How's this kit for tightening timings and possibly reaching 6400?

G.Skill Flare X5 Series 32GB (2 x 16GB) DDR5-6000 PC5-48000 CL32 Dual Channel Desktop Memory Kit F5-6000J3238F16GX2-FX5 - - Micro Center

Get it now! Designed for DDR5-enabled AMD AM5 platforms, Flare X5 series DDR5 memory is built for high performance in a low-profile design. Featuring AMD EXPO overclocking profiles for easy memory overclocking, users can easily enable overclocked performance by simply enabling EXPO in the BIOS...

www.microcenter.com

 

[Dyno]

ah that's a typo, I used 23H2 of course. Nice find, fixed in all affected reviews (everything after 14900KS). 24H2 isn't out yet, only a beta, and Microsoft's "RTM" bs isn't worth anything these days when so many things are still broken and you get major patches on day one.

The OS was fully updated in April 2024, and then locked down, so it stays constant for the lifetime of the platform, to keep things identical for all the ~45 CPUs tested

Once X3D, Arrow Lake, 24H2 are released I will spend a few weeks full time, retesting everything, also new app versions, new BIOS, new drivers, new games

What performance does the AMD AGESA PI 1.2.0.2 bring? Would it effect game at all? Did you guys come out with an article in regards to Windows 11 24H2 AMD patch yet?

 

[W1zzard]

What performance does the AMD AGESA PI 1.2.0.2 bring? Would it effect game at all? Did you guys come out with an article in regards to Windows 11 24H2 AMD patch yet?

You mean with 24H2 Beta? No, others tested it already, I focused on this article instead. Once 24H2 final is out we'll test it of course, and I'm sure there will be even more performance improvements. Next AGESA will have more improvements, too, so stay tuned

 

[Chrispy_]

? Are you serious ? go there with a scope and measure, than come back to myself and show me the nanoseconds it takes for a full cycle please. Than calculate.
Hint: frequency = 1 / time
Hint: Double Data RATE

You're confusing CLOCK frequency with TRANSFER frequency, which was the point I made really clear, I thought.

"Frequency" in DRAM is not a term uniquely used for the clocking operations. Yes. The clock pulse frequency or "clockspeed" of DDR is half that of the actual operating frequency, but this clock pulse by itself is merely a control signal used to sync the pacing and rate of the actual transfer frequency, which is the real number of times a second your DDR is sending or receiving data.

Think of it like humans walking or running, for example. All of the terminology is in steps/cadence/paces/strides and they are the most obvious, default way to talk about something that has an underlying frequency of HALF the steps/paces/strides per minute, The "clockspeed" of your legs is rate at which you swing a leg through one complete cycle - say from the point your heel strikes the ground until the point at which the same heel strikes the ground again - but for every single one of those, you take TWO paces. Nobody calls them half-paces or half-steps, yet that's what they are if you are incorrectly laser focused on the actual cycle rate of each leg.

 

[Evrsr]

Great article. It would be cool to see some tighter timings at 4800 and 5600, but it seems very few are selling those kinds of kits.
I guess it is not as bad and less problems with stability to be had.

See one 5600 CL36 but they really are few available.