When you FCLK your UCLK!

[mackerel]

To keep it short, I just noticed something on my 7800X3D + 2x 32GB 6000C30 ram system. Asus mobo using EXPO II. MCR also enabled in case that makes any difference.

6000C30: FCLK 2000, UCLK 1500
4800C40: FCLK 1800, UCLK 2400

I never paid attention to FCLK or UCLK before today. In short, is UCLK of 3000 "safe"? I'm about to try turning it on and see what happens. I never noticed it was running half UCLK up to now. Just turned on EXPO and forgot about it.

Edit: "safe" = high chance of working without needing to do further tweaks.

Edit 2: I eventually found the setting hidden away in several layers of submenus in the overclocking section of bios. Booted first time after. System hasn't crashed yet so I guess it is ok? Proper stability checks will follow later. I went straight to my Prime95 test, and that gave 4% improvement over 1500 UCLK.

 

[mackerel]

I'm learning so much about AMD ram configuration...
UCLK ratio is also found in timing screen of my mobo bios so no need to dig deep into OC settings area
New ram would not boot with MCR previously enabled. I waited... and waited... and waited. Nothing. Had to put in old ram, which booted quickly, turn off MCR, then the new ram would boot.

The 3000 vs 1500 UCLK thing:
With 1R modules, it seems to auto run UCLK 3000 for 6000 ram.
With 2R modules, it seems to auto run UCLK 1500 for 6000 ram.
With 2R modules, it seems to auto run UCLK 2900 for 5800 ram.
So the divider switch over point seems different depending on rank.

For my Prime95 8192k 8c8w ram benchmark:
6000C30 2R vs 1R: +7.4% (both 3000 UCLK)
4800C40 2R vs 1R: +9.2%
6000C30 2R, 3000 vs 1500 UCLK: +4.1%

It is conclusive that rank still matters like it did DDR4 era. I suspect the difference could be even greater if I had a faster compute CPU than 7800X3D which was bought primarily for gaming.

 

[EarthDog]

It is conclusive that rank still matters like it did DDR4 era.

In P95. What about in more common workloads? Rendering, Encoding, etc.?

 

[mackerel]

In P95. What about in more common workloads? Rendering, Encoding, etc.?

The question I was looking to answer, and have answered, is does 2R ram still improve performance over 1R at same speed and timings? The answer is yes. How much and where is secondary in scope, and I think I also need to focus more on the CPU itself. For example, I could run single channel ram to simulate a 16 core with only 8 cores.

I've done enough benchmarking for one day and will be busy for much of the next week on other things before I can return to more testing. The only other rank test I did is Y-cruncher Pi 1B gets around 2% at 4800C40. That UCLK was at 1500 threw a spanner in the works, and I'll need to rerun a full set at 3000 at some point.

In the testing I have identified many tests that show signs of ram impact. The Junkshop subtest in Blender does, whereas monster and classroom don't significantly. Cinebench 2024 is known to be ram affected. Older ones still aren't. Pov-ray is not ram impacted at all so I can ditch that from my test list. It is a short test so others might still use it for general CPU benches. Specific subtests in Geekbench are impacted to varying degrees but I need more data to get to grips with it. It is interesting I'm seeing that in single core as well as multi core, so that might be more from latency than bandwidth. 7-zip is another that is interesting.

I had forgotten how much of a time sink benchmarking is. Made worse in that sometimes differences don't make sense, so requires double checking if it is real or did something weird happen. And GPU testing, I often have to disregard the 1st run, which can be high or low depending on the use. High if GPU is cool so boosts more than after warming up, low if game doesn't precache so 1st run needed to load things.

One question answered, opening up many new questions.

Edit: this is the other testing I did today. 6000C30 2R, 2000 FCLK, 1500 UCLK, vs 4800C40 2R, 1800 FCLK, 2800 UCLK.
Blender 4.2.0
Monster +0.3%
Junkshop +2.2%
Classroom +0.2%
Overall +1.0%

Pov-ray 3.7: -0.4%

Cinebench
R15: +0.3%
R23: +0.0%
2024: +1.9%

Prime95 8192k FFT
8c1w: -0.2%
8c8w: +14%

7-Zip 24.08
Compression: +5.1%
Decompression: +0.9%

Geekbench 6.3.0:
Single core +2.2%
Multi core +5.6%

Time Spy CPU +3.6%

Everything was run 3 times, except Time Spy CPU which was run 6 times each as it had higher variation (the difference is bigger than variation), and Geekbench which was run twice.

 

[EarthDog]

For sure it's a time sink, lol.

So, yes is the answer, but 'resounding' doesn't come to mind looking at these data sets. Seven of them (a majority) are under 1% or actually slower.

 

[mackerel]

So, yes is the answer, but 'resounding' doesn't come to mind looking at these data sets. Seven of them (a majority) are under 1% or actually slower.

That testing is for 4800C40 2R vs 6000C30 2R, it is not 1R vs 2R test which is to be completed at some future time. Regardless, it helps indicate which workloads are at all sensitive to ram performance. It being weakly scaling on this test system doesn't matter. I probably have the worst modern system to do ram testing since I have a low core count and big L3 cache. One option I was proposing was to simulate the effects of a hypothetical 16 core system by running single channel, thus cutting bandwidth in half and doubling the load ratio. I'm still thinking of getting that 16 core Zen 5 at some point.

 

[Sologuy]

OP, have you run gaming benchmarks to see differences? Time spy is only one, real world game engines are another thing altogether. Productivity benchmarks are only half the equation so to speak. You have an X3D chip, large cache can hide stuff, but in non X3D chips, Running 1:1 with highest stable FCLK matters more with outright performance in 3D gaming, gotta get those 1% lows raised!

 

[mackerel]

Looking back, the way I started this thread doesn't make it clear. My primary goal from this testing was to answer one question: in DDR5 era does having 2R ram give a performance increase over 1R ram at same speed and timings? I have answered that with a conclusive yes. How much performance and where is secondary and something I'll look into, possibly late next week as my systems are busy on other things for now. I have probably about a day of testing to do. For my personal interests, testing gaming performance is a very low priority. The introduction of a GPU takes away from CPU quite a bit, especially as I don't have a 4090 lying around, and the tests are often quite time consuming. The main benefit of this testing is for ram heavy workloads that behave like Prime95 since neither x96 CPU provider is giving us affordable access to sufficient bandwidth to feed the cores. 3D cache is one partial solution.

 

[EarthDog]

I have answered that with a conclusive yes.

I'm not trying to be difficult, but it really seems to depend on the test. When over half the tests show less than 1% (margin of error) increase or the same but slower.... I'd have to add a caveat to 'a conclusive yes'. You're ready to stamp this as 'so', but the data we've seen tells me otherwise (at least on more common benchmarks/activities).

 

[mackerel]

I'm not trying to be difficult, but it really seems to depend on the test.

You're still missing the main point. I'm looking for a yes/no answer to rank effects on DDR5 since someone raised that question previously. That it does is now beyond question. It doesn't matter if it doesn't benefit every test, it only needs be seen in one to show it is present at all. Changing SSD is not going to improve Cinebench scores, but that doesn't mean it has no impact elsewhere.

Let's move on. "How much" will vary a lot with test and hardware and will be investigated separately at a later time. Also this is NOT a product level test, but lower level. My intention of testing wider is to identify potential software candidates which may aid in assessing this area in future e.g. when looking at other people's testing, I can focus on those tests that can give better indication. The testing is leaning towards where it matters, not where it doesn't.

I have to say again, the bulk test results I was showing earlier were varying speed/timings, not rank! That was a side result of the pre-testing, as I was going to compare both conditions. I have NOT performed the rank tests since the MCLK thing means I have to redo a lot of it. The limited spot checks I did do showed what I needed to see. Oh, the pre-testing was also interesting in that it showed how much each test varied with multiple runs. Some were far better than 1%, others were much bigger. That does need to be considered also when claiming a difference, but for what I needed to see the difference in perf was far greater than the variation.

The 4 sets of tests I wanted to do were:
A: JEDEC 2R
B: EXPO 2R
C: JEDEC 1R
D: EXPO 1R
My old ram allowed me to pre-test A and B, and I could compare B vs A, which is the bulk results shown earlier. My new ram allows me to test C and D, and eventually compare A vs C, and B vs D. This has not been done apart from spot checks.

Since the previous testing, there has been a set of Windows Updates as well as a GPU driver update. I might as well take my BIOS up to latest release before redoing the tests to make sure everything is as up to date as possible.

I'm doing my testing for my benefit. I'm sharing it in case it is of interest or might help others.


Edit:
I thought of a shorter parallel:
When X3D first came out, was it fast in gaming more often than not, right? But it wasn't fastest in every test vs non-3D. Does that invalidate 3D helping with speeds in some things? No. I see rank the same way. With DDR5, 2R will generally help vs same speed/timing 1R, as we have already seen in DDR4. When it helps, great. If it doesn't elsewhere, that doesn't negate the fact it does help where it does.

 

[EarthDog]

You're still missing the main point. I'm looking for a yes/no answer to rank effects on DDR5 since someone raised that question previously. That it does is now beyond question. It doesn't matter if it doesn't benefit every test,

But it does, though. If you don't do anything that shows a gain... there's no gain for you. And, in the majority of your testing, there wasn't any. Please understand I'm not trying to minimize your findings. You're right. When it works, it works to the tune of a few %.

I guess I'm taking exception to the fact that you seem to want to boil it down to yes or no and from your own testing, it's clearly not that. There are plenty of caveats to say 'yes'. If someone asked me that question, I'd say, "it depends on the test and what you consider an improvement". To just say yes without explanation leads you to people thinking an SSD improves CB times because 'it's faster'... it is faster than a HDD, no doubt, but having a faster HDD doesn't mean it improves X,Y, or Z, but could mean it improves A,B,C!

Im sure this feels pedantic, and for that I apologize. But seeing these tests, I can't walk away with it as just a plain old "yes" without explain the details that it's not close to being across the board, or significant improvements (within the scope of this testing). Again, not trying to take away from your findings at all... it's just how I'm reading the data.

I guess if you weren't showing rank in the test, then I'm really confused... Thread title is one thing, one data is one, another data is other... maybe I'm just toast after my job interview, LOL! I guess I'll pipe down until you test rank alone...?

 

[mackerel]

I guess if you weren't showing rank in the test, then I'm really confused... Thread title is one thing, one data is one, another data is other... maybe I'm just toast after my job interview, LOL! I guess I'll pipe down until you test rank alone...?

We are in the classic case of I'm talking one thing, you're talking another, hence we're not meeting up.

I started this thread to report that, during my testing, I found out my system defaulted to half MCLK when using EXPO with 6000 2R ram. This was unexpected and I never looked before. Testing since then has shown if I manually set 1:1 it still works fine and is what I'm using right now. There was a general small % improvement from that, but I have not listed the results here. Also the system does default to 1:1 with 6000 1R and 5800 2R, so it appears to be taking rank into consideration when setting that MCLK.

As I was replying to that, I did mention in passing I did a spot check on the 2R vs 1R testing that was my end goal, and that confirmed a difference was there. This was not meant to be my main post for that testing, since I have not completed it yet.

Also as mentioned before, I tested in two scenarios with 2R ram in preparation before receiving the 1R ram, and that was the testing I did report. I included it to illustrate what software might be influenced more or less by ram.

None of the testing or results should be a surprise to anyone as we've seen the same patterns with DDR4. I was verifying it was still doing it on DDR5 as there have been unsubstantiated claims otherwise.

Give me at least a week, I'll do my testing again with further system updates. I'll present the data then, whatever it is. This has given me some feedback on how to better present it to try and reduce misunderstanding. I am very specifically targeting some things and that could be made clearer.