Trident Z5 DDR5 7200mhz 2x16gb with Ryzen 7600x - timings and benchmark

[delvechio212]

After extensive testing i managed to set up timings to get best performance i could out of this ram kit.
G.SKILL Trident Z5 RGB Series CL34-45-45-115 1.40V F5-7200J3445G16GX2-TZ5RS
RYZEN 5 7600X @ 5.7GHZ
ASUS X570-P with 1654 bios – newest agesa 25/08/2023
Infinity fabric clock at 2100mhz
TRFC 480 @ 133,33 ns

These were best timings i could reach and still be stable in games and in overnight TestMem5.
If anyone has further suggestions let me know.. I will be testing further with increasing ram speed from here..
I also managed to do a 8000mhz stable ram frequency but timings were so horrible it was not even worth it.
https://hwbot.org/submission/5325923_delvechio212_memory_frequency_ddr5_sdram_4000_mhz
Memory Timings
CAS Latency (CL) 36T
RAS To CAS Delay (tRCD) 46T
RAS Precharge (tRP) 38T
RAS Active Time (tRAS) 36T
Row Cycle Time (tRC) 72T
Row Refresh Cycle Time (tRFC) 312T, 2x Fine: 192T
Command Rate (CR) 1T
RAS To RAS Delay (tRRD) Different Rank: 0T, Same Bank Group: 8T, Diff. Bank Group: 4T
Write Recovery Time (tWR) 48T
Read To Read Delay (tRTR) Different Rank: 12T, Different DIMM: 6T, Same Bank Group: 6T, Diff. Bank Group: 1T
Read To Write Delay (tRTW) 18T
Write To Read Delay (tWTR) 7T, Same Bank Group: 22T, Diff. Bank Group: 6T
Write To Write Delay (tWTW) Different Rank: 15T, Different DIMM: 7T, Same Bank Group: 6T, Diff. Bank Group: 1T
Read To Precharge Delay (tRTP) 12T
Four Activate Window Delay (tFAW) 20T
Write CAS Latency (tWCL) 34T
Write RAS To CAS Delay (tRCDW) 46T
Refresh Period (tREF) 65535T

 

[Woomack]

This RAM should work at CL32/34 but CL has to be equal or higher than wCL and you have to bump voltages. Typically, you set wCL -2, so when CL is 36, then wCL is 34. Safe way is to set it to the same value. I haven't seen significant differences at lower CL, so I see no point in pushing it at high voltages only to drop CL by one step.
I had no problem with tFAW at 16 on any Hynix A/M kit. Maybe it doesn't work in your case, I'm not sure why.

That AIDA64 is showing errors. There is no way you can get so high bandwidth at these settings. It sometimes happens. I get random 130-150GB/s. On the other hand, these settings should give you ~55ns latency.
At 8000MT/s timings are much worse and you can't make much about it. Somehow, it looks much better at 7800MT/s and it's not so far. 7600MT/s seems the best as it runs at not much worse timings than 7200MT/s.

I had no time to play with the whole timing table on AMD. However, tweaking RAM on AMD is pretty disappointing. I get almost the same results (out of synthetic benchmarks) at most settings. Every motherboard also couldn't work at more than 8000MT/s, so it's like 6200-6400 1:1 and then 7200-8000 1:2, when the average 6200 is as fast as 7600.

 

[Sologuy]

This RAM should work at CL32/34 but CL has to be equal or higher than wCL and you have to bump voltages. Typically, you set wCL -2, so when CL is 36, then wCL is 34. Safe way is to set it to the same value. I haven't seen significant differences at lower CL, so I see no point in pushing it at high voltages only to drop CL by one step.
I had no problem with tFAW at 16 on any Hynix A/M kit. Maybe it doesn't work in your case, I'm not sure why.

That AIDA64 is showing errors. There is no way you can get so high bandwidth at these settings. It sometimes happens. I get random 130-150GB/s. On the other hand, these settings should give you ~55ns latency.
At 8000MT/s timings are much worse and you can't make much about it. Somehow, it looks much better at 7800MT/s and it's not so far. 7600MT/s seems the best as it runs at not much worse timings than 7200MT/s.

I had no time to play with the whole timing table on AMD. However, tweaking RAM on AMD is pretty disappointing. I get almost the same results (out of synthetic benchmarks) at most settings. Every motherboard also couldn't work at more than 8000MT/s, so it's like 6200-6400 1:1 and then 7200-8000 1:2, when the average 6200 is as fast as 7600.

Interesting comments there Woomack, just wanted to clarify if you meant' wCL is tCWL (typo?)
With a single CCD chip, isn't 1:1 a better option for performance than 1:2? at least that is the consensus I've seen around the net on other forums so far with Zen 4.

I was looking at this kit as my next ram upgrade so when in 1:1 with up to 6600, one should get some pretty tight timings happening no doubt.

 

[Woomack]

Interesting comments there Woomack, just wanted to clarify if you meant' wCL is tCWL (typo?)
With a single CCD chip, isn't 1:1 a better option for performance than 1:2? at least that is the consensus I've seen around the net on other forums so far with Zen 4.

W as writing. Various motherboards have it called in a bit different way. The same as various brands call other timings in a different way. It means the same.

A 1:2 ratio on AM5 works for a few months, so there are barely any good comparisons. There is literally 1 memory series with an EXPO profile above 6400, and I think it was more by mistake, as it couldn't even be tested. However, the same as on AM4, you have to reach some frequency to match the performance.
In short, on AM4/non-APU, to match 3600 1:1, you had to set around 4600-4800 1:2. Non-APU AM4 chips could barely reach 5000, and most couldn't pass 4800. APU 4000/5000 series could go up to 4400+ 1:1, but the timings were not the best, so you could match it with 5200-5400 1:2 at relaxed timings. Still, not many AM4 motherboards support 4800+, so for daily/gaming PCs, the best is to set 3600-3800 1:1.
The same goes for DDR5. You have to pass some frequency to see a similar performance. In my tests, it was 7600 1:2 to match 6200 1:1, but it also depends on used timings, FCLK, and other things. Most motherboards will have problems with timings at 7200+.

There is also a performance difference between single and dual rank. On DDR4 it matters, on DDR5 it doesn't matter at all. This is why 2x24GB DDR5 kits have a point - they overclock easier than 2x32GB, and still give you a higher capacity than 2x16GB.

 

[delvechio212]

So if you check this picture


It shows 6400mhz 1:1 with lower timings and 7200mhz 1:2 with a lowest stable.
In pyprime it has really small difference.
Several time spy and firestrike runs gave me almost same results..
Also trying GDM on and off was such a small difference i had to do several runs and average them to notice a slight difference.. Run to run variance was greater than the actual frame / score difference.

So basicaly there is no difference for running pyprime at least in those two frequency options..

Also i want to add that my CPU is dual CCD version, with one CCD inactive.

 

[Woomack]

In theory, you should try for as high a memory copy as possible and the lowest latency. This is how the OS translates RAM performance, and somehow it works in daily usage and 3D benchmarks.
I'm not sure what I'm missing on these screenshots, but your memory read bandwidth is weirdly high, and the memory copy on the 7200 screenshot, too. On the other hand, I have about 55ns latency at not fully adjusted timings at @7600 and @8000. Here is one link. About the same was on G.Skill 2x24GB at 7200 and 7600. As I said, I didn't have much time for tests on AMD, and recently I'm mainly testing 24GB modules that require more relaxed timings.

 

[delvechio212]

Your picture shows 99.000 MBS mem read value for ryzen 9 7950x at 8000mhz , and i have 100.000 MBS at 7200 and slightly lower timing but still close to yours.
Which might be making sense as both your cpu and mine are DUAL CCD that have full benefit of both interconnects from core to memory..

But you should still have more or i should have less, its too close for difference in ram timing and frequency.

Also i have 5.75ghz locked max core frequency in hydra 1.4b


Is it possible that your 8000mz is having some issues , like if its not memory corruption of bits, but actual drops of full refresh/read cycles not being done or skipped?
Or that my aida64 is being retarded and lost the ability to count properly..

Also my attempt at 8000mhz is here

delvechio212`s Memory Frequency score: 4000 MHz with a DDR5 SDRAM

The DDR5 SDRAM @ 6000MHzscores getScoreFormatted in the Memory Frequency benchmark. delvechio212ranks #146 worldwide and #146 in the hardware class. Find out more at HWBOT.

hwbot.org

that was almost on the day the bios update with agesa came out.. it was unstable and i went back to my old settings..
think i might play around and see what could i get from 8000, if anything.

 

[Woomack]

Just some quick thoughts, if I remember it correctly.
I have +/- 2GB/s between my 7600, 7800 and 8000 results at not much different timings. I still have 8-9ns lower latency than your 7200 result and it's a lot considering timings.
Typically, 6200-6400 goes up to 90GB/s read/write and around 70GB/s copy ... not 110GB/s read. 7200-8000 is more like 90-100GB/s and 5-6ns worse (because of that 1:2 ratio) at more relaxed timings (65-70ns), that can be tightened to about 52-55ns. As I mentioned, 7600 is not far, even in synthetic results, from the 8000.
On the other hand, it's hard to stabilize any tighter timings on AMD. Sometimes it seems stable but crashes randomly after some time. In benchmarks it sometimes shows weird results. Each time I was testing something on AMD, I had to significantly reduce mulitple timings to keep stability. However, as I said, I didn't test enough to give any details.

AIDA64 benchmarks are multithreaded, it doesn't matter much how high is the CPU frequency, but more threads give higher bandwidth. It can be correct with the CCD too, as my single CCD Ryzen 7600 has about 20GB/s worse memory read and maybe 10GB/s worse memory copy. Memory write is not so bad.

I would try to get to 7600, stabilize it, and then play with timings, fclk and other options to get as much as it's possible. I expect it can be more for benchmarks than daily usage, but it's still good to start from any stable settings as some demanding benchmarks are not far from stability tests. Some even crash faster than stability tests because of different test patterns.

Once I finish upcoming reviews (as usual everything is delayed ), then I rebuild AMD test rig and try to push RAM some more.

 

[delvechio212]

6400
GDM enabled
tFAW 20
111 GBs
62.2 ns



6400
GDM disabled
tFAW 16
111 GBs
59,7 ns



6400
GDM disabled
tFAW 20
113 GBs
TRFC 380
59.9ns

 

[Woomack]

I'm not saying it's not correct, but these results on screenshots are all over the place. I mean on the previous screenshot there was 72GB/s memory copy and here is 116GB/s at almost the same settings. At these settings you should have lower latency. Again, I don't know what I'm missing here
I also wonder if motherboards may limit the bandwidth like it was on some previous generations. I saw it mainly on Intel motherboards. Maybe AMD messed up something with CPU batches as results that I see on the forums (also in some other threads) are sometimes hard to compare.
I have to find some time and switch motherboards in my ITX PC. Right now I have only ASRock B650E ITX free, which is fine, but acts weird at tighter timings. ASUS B650E-I would be much better for RAM OC.

Either way, I'm not sure if you can improve it more. It's pretty much maxed out and if you can stabilize it for 24/7 use then it would be great. I guess you can only try some more for 7600-8000 at tight timings, but I'm not sure if you see any better results.

Btw. tREFI doesn't work any higher? I see it's at 20k. Usually going for 64k+ lowers latency by 2-5ns.

 

[delvechio212]

TRFI is maxed at 65535 as a set value.
20k value is in nanoseconds.

 

[Woomack]

My bad, I really need to get glasses ...

 

[Sologuy]

Is Aida64 the only go to benchmark for comparing this phenomena?

 

[Woomack]

Is Aida64 the only go to benchmark for comparing this phenomena?

I had no time to test it. AIDA64 is generally the best benchmark for that, and usually there are no problems. There is clearly a significant difference between 1 CCD and 2 CCD, but I don't get why on two CPUs with 2 CCD, there are so different results. I haven't seen more detailed X3D CPU results. I'm still thinking of maybe buying 7800X3D, but soon, it will be Intel 14th gen CPU premiere, and I wish to get one too. My budget won't let me buy everything only for tests. I'm already spending much more than I was supposed to on hardware reviews.

Today, I started testing Dominator Titanium 2x24GB 7200 on the X670E Gene motherboard. At XMP settings, so 7200 CL36-46-46, with the 7950X CPU, I get 90/90/80 GB/s and 63ns latency. 24GB modules run at worse sub-timings than 16GB. At XMP with a 7600 memory clock, I get 92/91/83 GB/s and 58ns latency.
Now latency surprises me as when I was testing the G.Skill 2x24 7600 kit on earlier BIOS, then XMP was like 66-70ns. Clearly something has changed in BIOS, and this is version 1602, so from late August. I also see that ASUS has 8000 G.Skill kits on QVL right now. It still crashes at 8000 on my setup.

Edit:
8000 passed 50mins of the stability test but at 1.45V IMC voltage ... and then was a bluescreen. I remember 1.40V was enough for the same on older BIOS. 1.40 is actually as high as the motherboard suggests.

 

[Sologuy]

I had no time to test it. AIDA64 is generally the best benchmark for that, and usually there are no problems. There is clearly a significant difference between 1 CCD and 2 CCD, but I don't get why on two CPUs with 2 CCD, there are so different results. I haven't seen more detailed X3D CPU results. I'm still thinking of maybe buying 7800X3D, but soon, it will be Intel 14th gen CPU premiere, and I wish to get one too. My budget won't let me buy everything only for tests. I'm already spending much more than I was supposed to on hardware reviews.

Today, I started testing Dominator Titanium 2x24GB 7200 on the X670E Gene motherboard. At XMP settings, so 7200 CL36-46-46, with the 7950X CPU, I get 90/90/80 GB/s and 63ns latency. 24GB modules run at worse sub-timings than 16GB. At XMP with a 7600 memory clock, I get 92/91/83 GB/s and 58ns latency.
Now latency surprises me as when I was testing the G.Skill 2x24 7600 kit on earlier BIOS, then XMP was like 66-70ns. Clearly something has changed in BIOS, and this is version 1602, so from late August. I also see that ASUS has 8000 G.Skill kits on QVL right now. It still crashes at 8000 on my setup.

Edit:
8000 passed 50mins of the stability test but at 1.45V IMC voltage ... and then was a bluescreen. I remember 1.40V was enough for the same on older BIOS. 1.40 is actually as high as the motherboard suggests.

Yes, Aida64 is the dominant app for this type of testing, no doubt but that creates a kind of monopoly with this application & I don't think that's healthy for the industry. Passmark have a memory testing module in their performance test but I rarely see it on enthusiasts forums when comparing systems. So in effect, about 99% of enthusiasts are relying on a Finalwire app for their be all & end all judgement on RAM performance.

I know 24GB modules are kinda trendy now, but I see no use for them in gaming atm. Game engines still have a long way to go to utilize system ram to that extent ( that is more than 32GB if gaming is the single focus of the system) Perhaps in 2033?
Do you think the motherboard vendors are implementing agesa updates in different ways across each of their respective AM5 board ranges? Because it seems since this socket was launched, Gigabyte & Asrock have had the best track records for OC performance with their boards.

 

[Woomack]

AIDA64 Cache & Memory Benchmark is only a part of the whole AIDA64 software. There are so many things built-in (that most people never use) it's hard to call it only a benchmark. In general, it's more like a reporting tool and there are many competitive options on the market.
It's not their fault that most memory benchmarks are discontinued, cost more or focus only on few things, so are less popular.
Benchmarks like Geekbench seem popular too, but for most people it's a waste of money and its results are not clear enough or harder to compare. Recently you see press releases of various products where often are early Geekbench results ... and comments from users that who cares about Geekbench. This is only an example.

I'm using up to 35GB of RAM on my gaming PC. I already tested that multiple times and I get anywhere between 30 and 35 GB. Longer sessions in Diablo IV or Baldur's Gate 3 are using 14-16GB RAM +OS, web browser and the stuff in the background 10GB+. I'm usually keeping one more game minimized as the character does some stuff on its own, so +8GB.
I understand that most users don't need more than 32GB, but 16GB is getting not enough (on work laptop with only browser and some other office applications running I'm reaching 16GB). 32GB and 48GB memory kits often cost almost as much, so when you have a chance then why not to get a 48GB kit.

Until the AGESAs that let to set higher memory clocks at 1:2 ratio, all leading motherboard brands were providing exactly the same OC. All could run up to 6400-6600, as long as the used CPU was good enough. ASUS always have the best RAM and BIOS support.
Differences between various brands are mainly in memory training and how the motherboard acts during overclocking. For me, the last generation ASRock AMD motherboards are meh. They overclock well (still as well as other brands), but training takes ages and sometimes they're not passing initial tests or there are cold boot issues. Also everything I tested has annoying coil whine in idle. This is my experience and doesn't have to repeat on every motherboard. However, in the last few years, ASUS is always the best choice and next is usually MSI. Gigabyte improved RAM compatibility and overclocking in the last 2 generations, but some of their motherboards are still acting weird during memory training. ASRock, on the other hand, is worse in the last two generations than most people remember. On AMD it's not so visible, but on Intel chipsets, a typical ASUS, Gigabyte or MSI can make 7800+ RAM clock, while typical ASRock 7200. Typical Biostar, or other less popular brand, usually around 6400-6600.
AGESA is the same as it's provided by AMD. The difference is how BIOS is prepared and tuned for RAM support and overclocking in general. There can be other variables that affect the performance. As long as most motherboard performance comparisons in hardware reviews are pointless (+/- 1% differences in results), then sometimes we see that some models perform 5 or even 10% worse than expected.

 

[delvechio212]

I did some testing on 7600mhz and 1-1 , these are the stable timings and aida64.

If i go tFAW 16 or tRFC down to 560 it starts to show errors.

These are not final timings tho, i will continue to play today and see what i can get to be stable at 7600.

 

[Woomack]

I saw some guys having luck with higher VDD MISC at a higher RAM frequency, like 1.30-1.35V.
I finally set stable 8000 CL38-48-48 1.45V, but MC voltage is 1.4625V. It's a new memory kit so I started from DOCP tweaked (I'm not sure if you have it in your BIOS). Only main timings are manual and everything else at auto. 7800 was passing at XMP/DOCP and full auto. Problems started at 8000.
16GB modules run at a bit tighter timings. I guess you can stabilize yours at 7800-8000 CL34-36.

 

[delvechio212]

I did y-cruncher at 7600 with those timings.

delvechio212`s y-cruncher - Pi-2.5b score: 1min 12sec 977ms with a Ryzen 5 7600X

The Ryzen 5 7600X @ 5500MHzscores getScoreFormatted in the y-cruncher - Pi-2.5b benchmark. delvechio212ranks #10 worldwide and #1 in the hardware class. Find out more at HWBOT.

hwbot.org

Will be playing around for 8000 this weekend.
I do have DOCP I, DOCP II, DOCP Tweaked, and also tried with DDR5 Nitro options,

I tried doing primary timings and others auto and it would not post at 8000.
I had it at 8000 at one point but was not stable there at all..
Probably will need to play around with voltages and check buildzoid videos, might get a useful idea there.

 

[Sologuy]

AIDA64 Cache & Memory Benchmark is only a part of the whole AIDA64 software. There are so many things built-in (that most people never use) it's hard to call it only a benchmark. In general, it's more like a reporting tool and there are many competitive options on the market.
It's not their fault that most memory benchmarks are discontinued, cost more or focus only on few things, so are less popular.
Benchmarks like Geekbench seem popular too, but for most people it's a waste of money and its results are not clear enough or harder to compare. Recently you see press releases of various products where often are early Geekbench results ... and comments from users that who cares about Geekbench. This is only an example.

I'm using up to 35GB of RAM on my gaming PC. I already tested that multiple times and I get anywhere between 30 and 35 GB. Longer sessions in Diablo IV or Baldur's Gate 3 are using 14-16GB RAM +OS, web browser and the stuff in the background 10GB+. I'm usually keeping one more game minimized as the character does some stuff on its own, so +8GB.
I understand that most users don't need more than 32GB, but 16GB is getting not enough (on work laptop with only browser and some other office applications running I'm reaching 16GB). 32GB and 48GB memory kits often cost almost as much, so when you have a chance then why not to get a 48GB kit.

Until the AGESAs that let to set higher memory clocks at 1:2 ratio, all leading motherboard brands were providing exactly the same OC. All could run up to 6400-6600, as long as the used CPU was good enough. ASUS always have the best RAM and BIOS support.
Differences between various brands are mainly in memory training and how the motherboard acts during overclocking. For me, the last generation ASRock AMD motherboards are meh. They overclock well (still as well as other brands), but training takes ages and sometimes they're not passing initial tests or there are cold boot issues. Also everything I tested has annoying coil whine in idle. This is my experience and doesn't have to repeat on every motherboard. However, in the last few years, ASUS is always the best choice and next is usually MSI. Gigabyte improved RAM compatibility and overclocking in the last 2 generations, but some of their motherboards are still acting weird during memory training. ASRock, on the other hand, is worse in the last two generations than most people remember. On AMD it's not so visible, but on Intel chipsets, a typical ASUS, Gigabyte or MSI can make 7800+ RAM clock, while typical ASRock 7200. Typical Biostar, or other less popular brand, usually around 6400-6600.
AGESA is the same as it's provided by AMD. The difference is how BIOS is prepared and tuned for RAM support and overclocking in general. There can be other variables that affect the performance. As long as most motherboard performance comparisons in hardware reviews are pointless (+/- 1% differences in results), then sometimes we see that some models perform 5 or even 10% worse than expected.

Yes, I'm well aware of Aida64 being more of an all-in-one tool. I'm just glad there are substitutes on the market & a lot are free too. Only had a few fleeting moments of actually using Geekbench in the past, but not a fan of it. However it is what it is.

The comments & impressions I've gathered from places like here as one example, indicated Asus was not the best with Zen 4. Many users of their boards were not happy with bios support & agesa updates, however that was several months back so perhaps things have changed now.
I have an MSI B650 Edge WiFi Board but was not impressed with its bios support for Zen 4. GDM was not even available a few weeks back with the latest bios then, a perusal of their support site for discussions about their motherboards reveals it's not uncommon for end users to have issues with their line of Zen 4 boards & bios, however again, perhaps this has changed now as that was several weeks back when I looked there.
The memory training time issue has been resolved with my Asrock board - at least with the latest bios release. However, the PBO customised curve optimizer settings have broken & discovered this the hard way via windows corruption... getting a bit off topic here, but for me from now on, I'm sticking only with official bios releases & no beta versions ever again with Zen 4 on any motherboard I have for them. So I reverted back to the 1.28 bios for the time being. Just sick n' tired of vendors fixing some things, but breaking others with bios releases. I don't know if when or ever this will settle down once Zen 5 is released.