DDR5 Memory Performance Scaling with AMD Zen 5

[W1zzard]

Our AMD Zen 5 Memory Scaling Review examines six different DDR5 memory speeds, including DDR5-8000, DDR5-6400 1:1, DDR5-6400 CL28 and others. We looked into how these speeds impact performance in gaming at four resolutions and a wide range of application workloads.

Show full review

 

[ir_cow]

Just to help people understand the MHz thing. DDR stands for Double Data Rate. Data is sent on the raise and fall of the clock cycle making the actual operating clock speed differ from the effective clock speed.

For example take DDR5-4800. It is effectively running at 4800 MHz, but actually at 2400 MHz. Another way of saying this is 4800 MT/s (MT = Mega Transfers). For many years now marketing has gotten in the way of correct information. Memory is no different. There is no 8000 MHz memory (yet).

 

[Awwwyeahhhbaby]

As expected 6000-8000 is a nothing burger, most averages show 1-2% variance which is margin of error and none of the speeds tested consistently took top spots. Best value for the money is still DDR5-6000 low latency and manually tune sub-timings.

I'd avoid 4800 and 5600 but most of us have been saying that since 7000 series.

 

[AtiAmd]

Well done Wizzard! Loads of testing and a lot of interesting results ☺
Yep ,we only need 6000 to 6400 and we should never over pay for something we don't really need.

 

[soulrider4ever]

Great article, hope to see a follow up to the "new" X870E variants to see if those mobo's have any noteable improvements in this regard.

 

[low]

thanks

 

[Chrispy_]

Three things:

1. Thanks for this, I've been waiting for Zen5 scaling answers and now I have them.
   
   
2. 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.
   
   
3. Good news for most of us, as DDR5-6000 CL30 EXPO kits have been the very affordable speed/price sweet spot for the last year or so and it seems there's very little point in paying more than that.

 

[kiakk]

My 10600K uncore idle ~3W (DDR4 3600), and
my 5600X IO idle ~19W (DDR4 3733).
As I see nothing changes after years.
I do not understand why AMD do not take care of idle/light load power management.
(I would ask AMD, does Global warming exist or NOT, WTF, are you crazy??? Every power saving matter.)

 

[ir_cow]

The ultimate performance mode would be 6400 MT/s with FLCK of 2133. Next change the CAS to the lowest you can go and adjust the tRFC, tFAW, tRRD_sg, tRDRD_sg, tWRWR_sg, tRFC and tREFi. Even just tweaking a 6000 kit with FLCK 2000 will beat out EXPO/XMP 6400/2000 (stock default).

 

[KarymidoN]

amazing job @W1zzard, thank you.

i have a question, is the controller in the 8000X processors any different from the 9000 ones? i've seen different memory behavior from 7000 / 8000 / 9000 series, from 7000 to 9000 it doenst appeared to have any improvements, but 8000 series APUs seem to have a better Memory controller, anyone knows why?

 

[redzo]

My 10600K uncore idle ~3W (DDR4 3600), and
my 5600X IO idle ~19W (DDR4 3733).
As I see nothing changes after years.
I do not understand why AMD do not take care of idle/light load power management.
(I would ask AMD, does Global warming exist or NOT, WTF, are you crazy??? Every power saving matter.)

That's one of the few sacrifices of the MCM design approach and it is looking like it is here to stay forever.

 

[Neo_Morpheus]

My 10600K uncore idle ~3W (DDR4 3600), and
my 5600X IO idle ~19W (DDR4 3733).
As I see nothing changes after years.
I do not understand why AMD do not take care of idle/light load power management.
(I would ask AMD, does Global warming exist or NOT, WTF, are you crazy??? Every power saving matter.)

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

Personally, an idle cpu will most likely be with the whole system in sleep mode.

 

[Chrispy_]

That's one of the few sacrifices of the MCM design approach and it is looking like it is here to stay forever.

It's a perfectly acceptable sacrifice. Without it, doubling the core count would quadruple the price we'd pay for them.

If anyone is worried about the extra 16W at idle, don't be. There are bigger problems that the human race needs to address before idle power consumption of a PC is a concern.

Popular EVs like the Tesla model X are 330Wh/mile if driven carefully, rather than with a heavy right foot - So a gentle, efficiently-driven commute to work 10 miles away and back is almost 7KWh, which will power the 16W of IO die of a Ryzen for over 400 hours, which is several months of PC usage for most people.

Want to save the planet? Stop eating meat, cycle or walk instead of moving your 50-100kg self around in a 2500kg chunk of steel, and don't have too many offspring.

 

[Dr. Dro]

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.

 

[Nhonho]

DDR5 6000 memory with CL 28 is very rare to find. I didn't even know it existed. You should have used a DDR5 6000 with CL 32 to 36, which are the most common, so that the test would show results closer to what people will have in their PCs.

 

[TristanX]

AMD must fix read speed (some 60GB/s), to be equal to write speed (almost 90 GB/s), for 1CCD CPUs

 

[SpaceSquirrel]

Have you tried a 7800X3D @ 4800 with tuned timings ( Buildzoid's Hynix A-Die EZ-timings guide ) at default voltages?

I did that by accident a while back and the results floored me.

I barely saw any performance regression from ez-tuned 6000 ( perhaps 1-2% ), what's more? It was FASTER than just enabling EXPO without tuning the RAM and I also saw a very notable decrease in power consumption both for the RAM sticks and the CPU.

Once I stumbled upon that, I never went back. I even undervolted VSoC to 0.85v for even more power savings, the 7800X3D now idles around 12W ( instead of 25+ @6000 ). Ran multiple RAM stability tests for weeks, never had an error.

I'd be curious if finally someone like you would experiment with that and "peer-review" my amateur findings.

It's with a G.Skill Trident Neo Z5 CL30 2x32GB kit btw ( F5-6000J3040G32G ). But I guess it would work even better with a CL28 kit.

Here's the ZenTimings screenshot for the timings and voltages running on the kit:

https://imgur.com/IC9l07w

Any chances of such an inquiry? Or perhaps with the 9800X3D once it hits the market?

 

[Pepamami]

have u also tried tuned memory? With less abysmal tRFS''s, tREFI, tRAS, tRS and etc numbers. That ram makers put unnecessary high.

 

[TomWeng]

Save users more expenses.

 

[xSneak]

Thanks for the review. It seems like 6000 Mhz is still really all you need. How much of a factor is the mobo in achieving 8000 mhz ? i see you guys are using a crosshair board.

 

[ir_cow]

have u also tried tuned memory? With less abysmal tRFS''s, tREFI, tRAS, tRS and etc numbers. That ram makers put unnecessary high.

They follow JEDEC so it can handle 85c for operational temperatures. Though this isn't always the case that higher speeds can handle 85c. Still it is "high" for a reason.

Thanks for the review. It seems like 6000 Mhz is still really all you need. How much of a factor is the mobo in achieving 8000 mhz ? i see you guys are using a crosshair board.

you can count on one hand the amount of motherboards that can do 8000 without crazy voltages. X870 will be better because vendors came up with some better layering and traces.

 

[CheapMeat]

Just wanted to say thank you for the time and effort on doing all the comparisons. This is why I love dedicated hardware enthusiast websites.

I wonder if the 9950X3D and 9900X3D with an X870E board make any significant difference in the comparisons. I think might finally upgrade from my X99 workstation.

 

[Daven]

I can't tell from the Test Setup notes but are these tests run using local admin mode? Or is this fixed with the latest Win 11 update? I remember reading something about slower AMD CPU performance regarding local vs 'run as admin' modes.

At any rate, great and insightful review W1zzard!

 

[Dr. Dro]

DDR5 6000 memory with CL 28 is very rare to find. I didn't even know it existed. You should have used a DDR5 6000 with CL 32 to 36, which are the most common, so that the test would show results closer to what people will have in their PCs.

6000 CL28 is the new "gold standard" for Zen 5, it's a very low latency kit. Not a budget-friendly option, nor the most common, but as DDR5 advances, should become more commonplace

I can't tell from the Test Setup notes but are these tests run using local admin mode? Or is this fixed with the latest Win 11 update? I remember reading something about slower AMD CPU performance regarding local vs 'run as admin' modes.

At any rate, great and insightful review W1zzard!

Issue's been supposedly fixed for some time now

 

[Deleted member 223428]

6000 CL28 is the new "gold standard" for Zen 5, it's a very low latency kit. Not a budget-friendly option, nor the most common, but as DDR5 advances, should become more commonplace

Yeah, the pre-binned kits are about 50% more than 6000/30. You probably could get there on a regular set of A-Die with some FAFO.