Which RAM to choose for AMD 7950X3D and keep a 1:1 ratio

[Chrispy_]

It sounds like the best chance of a headache-free option is a 2x32GB kit of 5600-CL36 EXPO.
No, it's not the fastest thing you can possibly run, but it'll still be quick and it should at least work without needing to know what you're doing with manual primary and secondary timings.

 

[curtisdodo]

Hello, following the advice of a Reddit member I'm going to go with the KF560C36BBEK2-64, so 2X32GB DDR5 6000 CL 36 38 38 1.35V, because they are the best hynix A-die chips with which the weakness and potential are better. I play in 4K, so I should have good performance.

 

[Klemc]

IDK if burned 7000 AMD had all EXPO RAM, mine is XMP 6000-32 and i didn't had the problem, all profiles work, 15 sec for training, whatever i disabled XMP i hate waiting training time.

 

[dgianstefani]

Hello, following the advice of a Reddit member I'm going to go with the KF560C36BBEK2-64, so 2X32GB DDR5 6000 CL 36 38 38 1.35V, because they are the best hynix A-die chips with which the weakness and potential are better. I play in 4K, so I should have good performance.

At 4K you wouldn't notice much of a difference with RAM configurations, because you are almost never going to be CPU limited.

Ideal ram tunes are for players who seek high frame rates and are thus CPU limited. 4K is GPU limited even with a 4090.

 

[ir_cow]

First problem, you took the advice of reddit .

 

[dgianstefani]

First problem, you took the advice of reddit .

To be fair A die is good ram.

 

[ir_cow]

To be fair A die is good ram.

True. The op lucked out.

 

[davidm71]

Just wanted to follow up and take a survey of how is it going with 2 x 32gb or 64gb ram setups? Wonder what vSoc you settled on?

I was on 1.24v for vsoc and all of a sudden had a random bsod from a bugcheck error. Did an SFC /scannow and found errors which I repaired.

But it seems that these 2 x 32gb setups need more voltage or lax the hell out of 2nd 3rd zoid timings perhaps?

Thanks

 

[The Von Matrices]

Just wanted to follow up and take a survey of how is it going with 2 x 32gb or 64gb ram setups? Wonder what vSoc you settled on?

I was on 1.24v for vsoc and all of a sudden had a random bsod from a bugcheck error. Did an SFC /scannow and found errors which I repaired.

But it seems that these 2 x 32gb setups need more voltage or lax the hell out of 2nd 3rd zoid timings perhaps?

Thanks

I've since upgraded to 2 x 48GB DIMMs but I'm using the exact same timings as the 2 x 32GB setup I had previously. I'm still using the same vSOC at 1.24V.

I found my minimum vSOC by starting at 1.3V and then dropping it until it could no longer pass two rounds of stress testing with memtest86 (the free version). I did add a fan get some airflow over the DIMMs since I am using a water block and the CPU socket doesn't get much airflow, and I think that hot DDR5 DIMMs may be an underreported cause of many crashes.

I haven't had a single memory-related crash in the entire year after I tuned the system and confirmed it passed memtest86.

 

[Sarajiel]

Wonder what vSoc you settled on?

Most BIOSes with somewhat current AGESA versions seem to default to Vsoc at around 1.2V +/- 50mV for 6000MT/s.
If you manually tune Vsoc, people tend to aim for 1.15V for tight timings and high FCLK nowadays.

Personally, I use 1.2V Vsoc for my 96GB setups at 6200 MT/s on a couple of Gigabyte boards.

I did add a fan get some airflow over the DIMMs since I am using a water block and the CPU socket doesn't get much airflow, and I think that hot DDR5 DIMMs may be an underreported cause of many crashes.

Not to be a smart ass here, but looking at your tREFI, you probably don't need a fan for your setup, because you refresh more often than JEDEC would require for your memory speed, aka 3.9µs. Keep in mind that JEDEC specs are typically intended for a maximum temperature of 80°C to 85°C. At the moment, you spend ~5.5% on refreshing your memory instead of the ~7.5% expected for your memory size by JEDEC spec.
The goal of memory overclocking would be to push this refresh time below 1% via adjusting tRFC & tREFI on AM5. Since this is a G.Skill kit, tRFC won't go much lower, but tREFI should easily go over 45000 cycles, ideally one wants to run this at the maximum of 65535 if thermally possible. Also, tRCD and tRP should work at 37 clock cycles, since you underclock your DDR5-6400 sticks anyway. For tRAS, I typically run low 30s on the same memory kit at 6200MT/s.

Generally speaking, you are right that most folks who overclock their memory ignore the heat on DDR5. For a water-cooled setup like yours, you could include a memory cooler in the loop. For example, Alphacool offers specific coolers for DDR5 with a matching water block for less than €100. However, strapping a 80mm or a 92mm fan on the back of your GPU, or using a specific memory cooler like the Jonsbo NF-1 is typically a bit cheaper, e.g. €5 to €25.

For thermal stressing memory sticks, one should run Prime95 (large FFTs) for about 24h to 36h while the GPU runs FurMark at the same time, at least if one uses a case.

 

[ir_cow]

Not to be a smart ass here, but looking at your tREFI, you probably don't need a fan for your setup, because you refresh more often than JEDEC would require for your memory speed, aka 3.9µs. Keep in mind that JEDEC specs are typically intended for a maximum temperature of 80°C to 85°C. At the moment, you spend ~5.5% on refreshing your memory instead of the ~7.5% expected for your memory size by JEDEC spec.
The goal of memory overclocking would be to push this refresh time below 1% via adjusting tRFC & tREFI on AM5. Since this is a G.Skill kit, tRFC won't go much lower, but tREFI should easily go over 45000 cycles, ideally one wants to run this at the maximum of 65535 if thermally possible. Also, tRCD and tRP should work at 37 clock cycles, since you underclock your DDR5-6400 sticks anyway. For tRAS, I typically run low 30s on the same memory kit at 6200MT/s.

Generally speaking, you are right that most folks who overclock their memory ignore the heat on DDR5. For a water-cooled setup like yours, you could include a memory cooler in the loop. For example, Alphacool offers specific coolers for DDR5 with a matching water block for less than €100. However, strapping a 80mm or a 92mm fan on the back of your GPU, or using a specific memory cooler like the Jonsbo NF-1 is typically a bit cheaper, e.g. €5 to €25.

I would agree that unless you are raising the tREFI a lot (65K), chances are you don't need a direct fan. Unless maybe if you are also going for 1.5+ voltage could be a valid reason to as well. I also found tRFC can become unstable if the temp goes up enough (see below). After all this is the refresh time (cycles) for the banks. They are ever so slightly thermally driven. So that last 25 might need below 45c. It also from observation the lowest value is kinda based on IC quality and of course IC brand / Revision.

For thermal stressing memory sticks, one should run Prime95 (large FFTs) for about 24h to 36h while the GPU runs FurMark at the same time, at least if one uses a case.

I don't know about this one. Memtest5 and Karhu will blast it way more. Prime95 will cook the CPU more than anything and FurMark is just for the GPU. Both together will create a lot of heat which will have to exit the case somehow. Good way to see if your PSU can handle the load . I will say that Furmark or Kombuster or even looping timespy is good to see if your memory will "overheat" from the GPU dumping hot air on it. NVIDIA 30 and 40 series rear fan points directly at my ram, I had to put a shim in-between to prevent crashes. Now its not a problem for me anymore with watercooling.

 

[Sarajiel]

I don't know about this one. Memtest5 and Karhu will blast it way more. Prime95 will cook the CPU more than anything and FurMark is just for the GPU. Both together will create a lot of heat which will have to exit the case somehow. Good way to see if your PSU can handle the load . I will say that Furmark or Kombuster or even looping timespy is good to see if your memory will "overheat" from the GPU dumping hot air on it. NVIDIA 30 and 40 series rear fan points directly at my ram, I had to put a shim in-between to prevent crashes. Now its not a problem for me anymore with watercooling.

I said Prime95 large FFTs. If you just loop through all sizes, the overall thermal stress isn't really that high. In theory, you could even try to find the worst size for your system, but I'm usually too lazy to do that.
From my personal experience with all kinds of 16-core Ryzens on 64GB to 192GB setups, this will run the memory sticks about 4°C to 7°C higher(*) than pretty much all other memory tests I've tried so far, regardless if they are pattern tests or memory-bound computational tests. However, I only extensively test daily setups in closed cases, not open benchtables.
Again, from my own experience, adding FurMark for increased thermal stress into the mix is definitely a good idea for air-cooled GPUs. However, since folks these days even dump hot air from their custom loops into their cases, because of the looks of their system, I'd say it might be a consideration for some water-cooled GPUs as well.
I could imagine that Prime95 might give different thermal results with LGA1700 CPUs since it uses a different code path, but I guess, I luckily skipped those?

What kind of case were you using? My GALAX 4090 can run at up to ~517W TGP and a 80mm Noctua fan on its back directed at the memory sticks is enough to cool the 96GB kit at 6200MT/s with tight timings. Maybe my good old Meshify 2 with a couple of 140mm Noctuas is still a champ when it comes to cooling?


(*) note: I don't have a fancy thermal imaging camera, therefore my temp measurements might not be too accurate, but a difference in temps was reproducible between different RTX 40 cards in different cases and different RAM configs.

 

[The Von Matrices]

Not to be a smart ass here, but looking at your tREFI, you probably don't need a fan for your setup, because you refresh more often than JEDEC would require for your memory speed, aka 3.9µs. Keep in mind that JEDEC specs are typically intended for a maximum temperature of 80°C to 85°C. At the moment, you spend ~5.5% on refreshing your memory instead of the ~7.5% expected for your memory size by JEDEC spec.
The goal of memory overclocking would be to push this refresh time below 1% via adjusting tRFC & tREFI on AM5. Since this is a G.Skill kit, tRFC won't go much lower, but tREFI should easily go over 45000 cycles, ideally one wants to run this at the maximum of 65535 if thermally possible. Also, tRCD and tRP should work at 37 clock cycles, since you underclock your DDR5-6400 sticks anyway. For tRAS, I typically run low 30s on the same memory kit at 6200MT/s.

Generally speaking, you are right that most folks who overclock their memory ignore the heat on DDR5. For a water-cooled setup like yours, you could include a memory cooler in the loop. For example, Alphacool offers specific coolers for DDR5 with a matching water block for less than €100. However, strapping a 80mm or a 92mm fan on the back of your GPU, or using a specific memory cooler like the Jonsbo NF-1 is typically a bit cheaper, e.g. €5 to €25.

Thanks for the feedback - no sarcasm detected. I'll try changing it and would be happy if you could provide any other timing tips.

Regarding memory cooling, my system is unique. I have a water block on the CPU and on my GPU, and the radiator is completely outside the case. Therefore, there is almost no airflow on the DIMMs. I have some fans pointed on the hard drives and some low-speed exhaust fans (speed proportional to VRM temps) but the case is barely moving any air. The DIMM fan is absolutely necessary in my system because otherwise they are effectively passively cooled.

 

[A Computer Guy]

I'm doing 1:1 with my kit, but at 5600. Can't say it's bothered me much.

Edit: Not x3d though.

At 1:1 is that 2800 infinity fabric ?

 

[Sarajiel]

Thanks for the feedback - no sarcasm detected. I'll try changing it and would be happy if you could provide any other timing tips.

You could use the tRP + tRAS = tRC formula for the delay timings. That allows you to tune tRP first and tRAS after without bothering about tRC. If the formula is ignored, the delay timing on the smaller side of the equation will be basically ignored by the memory controller, at the moment that would be tRC.
You could also try to tighten parts of the multi-rank timings for a small gain: tRDRDSD should be able to do 6, tRDRDDD isn't used in your setup, but some folks always set it to the same value as tRDRDSD, because some BIOSes for older Intel CPUs had them the wrong way round.

Other than that, your timings look pretty similar to my version of that memory kit.
That said, don't forget to run stability tests. Personally, I like to run a mix of at least one computational test (e.g. y-cruncher or Linpack Xtreme), one pattern tester (e.g. TestMem5, GSAT, Karhu, etc.), and one thermal test (Prime95 large FFTs or whatever makes your sticks even hotter).

the radiator is completely outside the case

That's quite the retro setup You definitely have to use decent fans for your memory sticks.

 

[R-T-B]

At 1:1 is that 2800 infinity fabric ?

Nah I meant UCLK and MCLK synced. Was not considering IF.

 

[KLMR]

I've got dual rank (64GB) 6000 CL36 running at 5800 CL30-36-etc. not the best OC in the world, but machine works for weeks at 100%.
I've tried so hard to lower those voltages...

My 2 cents:
- Secondary timings gave me "a lot" of performance boost
- Indeed I found cooling ram with a fan helped a lot in stability (WC block on CPU).
- The route to go is target lower memory Hz to try "1:1" and then tight timings as much as possible.

-