RVIE: Memory Write Much Worse Than Read and Copy

[gridironcpj]

I recently noticed that my memory write speed is much lower than my read or copy speeds when tested in AIDA64's cache and memory benchmark. For reference, I am using a Rampage VI Extreme and a 7920X. The memory used in the results below is a Trident Z RGB 32GB C16 3600 kit manually overclocked to 3800MHz with tighter secondary and tertiary timings.



As you can see, the write speed is significantly lower than the read and copy speeds. Every result I've seen with Skylake-X CPUs show much better results. Based on similar memory overclocks, I should at least be over 95K in write speed.

So far, to diagnose the issue, I tried running a different BIOS (1003 instead of 1004) and that did not make a difference. I then put my old Corsair 32GB C16 3200 kit that I used for my X99 build. I manually set the timings to be more or less the same as the my manual settings for the Trident Z kit, but kept the speed at 3200MHz. Here were those results:



So despite the lower memory speed, the Corsair kit is able to hit a slightly higher memory write speed. Just to be on the safe side, I returned my Trident Z RGB kit. I will probably replace it with the same kit, although I'm open to suggestions. Even the Corsair kit is showing a big gap between read and write speeds. My worry is that my motherboard or CPU are defective...? I'll note that I have not had any odd benchmark scores anywhere else except in this write speed portion of AIDA64's cache and memory benchmark. My system performs just fine with no issues.

I'm no expert on the matter, but could it be something off with my CPU's IMC? Or maybe my motherboard has a fault dimm slot or two? I'd love some suggestions on the matter since this small issue is sort of driving me crazy.

[CSN7]

Hah! Glad that someone brings this up! I noticed the same behavior. I was freaking out about it, too. I even changed from an EVGA X299 board to an APEX because I thought the EVGA board was the cause. Nope same situation the write speeds seem to be stuck. You will notice it can mainly be affect by Mesh frequency.

One thing we have in common is the 7920X.

Either the smallest of the HCC die chips has been somewhat cut down on the memory controller by intel hence there is not a single review unit send out by intel so there are literally no reviews on the 7920X out there so no one notices it's shortcomings OR Aida64 just doesn't monitor the write speed correctly on this chip. Because what leads me to this conclusion is that I tried other benchmarks and they actually report a write speed that is somewhat higher than read and copy, more like it's supposed to be.

For reference this is a Corsair Dominator Torque SE 3200Mhz (I have tightened the second & third timings a lot and Mesh runs @3200, but same overall scenario):

[restsugavan]

Hah! Glad that someone brings this up! I noticed the same behavior. I was freaking out about it, too. I even changed from an EVGA X299 board to an APEX because I thought the EVGA board was the cause. Nope same situation the write speeds seem to be stuck. You will notice it can mainly be affect by Mesh frequency.

One thing we have in common is the 7920X.

Either the smallest of the HCC die chips has been somewhat cut down on the memory controller by intel hence there is not a single review unit send out by intel so there are literally no reviews on the 7920X out there so no one notices it's shortcomings OR Aida64 just doesn't monitor the write speed correctly on this chip. Because what leads me to this conclusion is that I tried other benchmarks and they actually report a write speed that is somewhat higher than read and copy, more like it's supposed to be.

For reference this is a Corsair Dominator Torque SE 3200Mhz (I have tightened the second & third timings a lot and Mesh runs @3200, but same overall scenario):

Due Skylake SP Core microarchitect L1 cache data cache are 512-bit wide with allow 2 x 64Byte load and 1 x 64Byte store.
Thus almost case Read Speed much faster than write.

[CSN7]

Okay makes sense, but still I've seen screens of 7960X / 7980XE that show normal write speeds. So something is slightly off.

[Menthol]

Have either of you tried one of the memory presets in the memory timing section of the bios? I don't know what presets are available for the R6E but the R6A has a 1.4v 4000mhz profile that works really well. I see some guys reporting it works for them using Gskill 3200mhz 14-14-14 kits and 3600mhz 16-16-16 kits, I am using it with 2 kits of 3600 15-15-15 at the moment, these type of bios presets is one of the reasons I like to purchase the top boards, it can help the experienced like myself. Even if you can't run at 4000mhz you can still select the profile and then set a lower speed and be in the ballpark so to speak.
What cache speed are you running?

[gridironcpj]

Hah! Glad that someone brings this up! I noticed the same behavior. I was freaking out about it, too. I even changed from an EVGA X299 board to an APEX because I thought the EVGA board was the cause. Nope same situation the write speeds seem to be stuck. You will notice it can mainly be affect by Mesh frequency.

One thing we have in common is the 7920X.

Either the smallest of the HCC die chips has been somewhat cut down on the memory controller by intel hence there is not a single review unit send out by intel so there are literally no reviews on the 7920X out there so no one notices it's shortcomings OR Aida64 just doesn't monitor the write speed correctly on this chip. Because what leads me to this conclusion is that I tried other benchmarks and they actually report a write speed that is somewhat higher than read and copy, more like it's supposed to be.

For reference this is a Corsair Dominator Torque SE 3200Mhz (I have tightened the second & third timings a lot and Mesh runs @3200, but same overall scenario):

Well this is both a relief and a slight worry. It's a relief in that my hardware probably isn't defective then. It's a slight worry because I've seen the 7820X and 7900X with better memory write results, so it would be unfortunate if the 7920X turned out to have the worst memory controller. I'm hoping it's just AIDA64 like you are thinking. What other benchmarks have you used for your memory?

I was debating between the 7920X and 7960X and I chose the 7920X due to its decent pricing per core versus the 7960X and I reasoned that the only difference in performance would be attributed to the core count difference.

[gridironcpj]

Have either of you tried one of the memory presets in the memory timing section of the bios? I don't know what presets are available for the R6E but the R6A has a 1.4v 4000mhz profile that works really well. I see some guys reporting it works for them using Gskill 3200mhz 14-14-14 kits and 3600mhz 16-16-16 kits, I am using it with 2 kits of 3600 15-15-15 at the moment, these type of bios presets is one of the reasons I like to purchase the top boards, it can help the experienced like myself. Even if you can't run at 4000mhz you can still select the profile and then set a lower speed and be in the ballpark so to speak.
What cache speed are you running?

I have the Extreme, which turns out to be a worse board for memory overclocking (go figure). I had so many issues trying to stabilize 4000MHz with the C16 3600MHz Trident Z RGB kit I had. I was able to pass 1000% of HCI only for it to fail before 50% the next day. Again, it could be the 7920X's memory controller somehow being worse than every other Skylake-X CPU or it could be the Extreme simply not being so Extreme anymore like it was in previous generations. I had to settle for a 3800MHz memory OC with good secondary and tertiaries.

My cache is at 3100MHz. I tried 3200, but I kept getting BSODs at the Windows login screen. This was the case even at 1.2V on the cache. I only attempted it with the 3800MHz memory OC though, so maybe a lower memory OC would not produce the BSODs. The voltage increase from 3100MHz to 3200MHz must be absurd for me. I can do 3000MHz at 1.05V and 3100MHz at 1.1V.

[CSN7]

Have either of you tried one of the memory presets in the memory timing section of the bios? I don't know what presets are available for the R6E but the R6A has a 1.4v 4000mhz profile that works really well. I see some guys reporting it works for them using Gskill 3200mhz 14-14-14 kits and 3600mhz 16-16-16 kits, I am using it with 2 kits of 3600 15-15-15 at the moment, these type of bios presets is one of the reasons I like to purchase the top boards, it can help the experienced like myself. Even if you can't run at 4000mhz you can still select the profile and then set a lower speed and be in the ballpark so to speak.
What cache speed are you running?

@Menthol yes I tried the preset to get an idea first but they are quiet a bit slower than my current timings, while still maintaining 1.35 vDimm. I could tighten the primary timings by just adding volts, but it doesn't offer much performance honestly.

For comparison:
@1.35 vDimm
0.875v SA
1.125v IO


And Mesh is @3200 anything above gets really difficult on the bigger chips unfortunately. Could pass XTU with 3300Mhz Mesh but that needs 1.32 IO and 1.34 vMesh so not suitable for 24/7 http://hwbot.org/submission/3754653_

[CSN7]

I have the Extreme, which turns out to be a worse board for memory overclocking (go figure). I had so many issues trying to stabilize 4000MHz with the C16 3600MHz Trident Z RGB kit I had. I was able to pass 1000% of HCI only for it to fail before 50% the next day. Again, it could be the 7920X's memory controller somehow being worse than every other Skylake-X CPU or it could be the Extreme simply not being so Extreme anymore like it was in previous generations. I had to settle for a 3800MHz memory OC with good secondary and tertiaries.

My cache is at 3100MHz. I tried 3200, but I kept getting BSODs at the Windows login screen. This was the case even at 1.2V on the cache. I only attempted it with the 3800MHz memory OC though, so maybe a lower memory OC would not produce the BSODs. The voltage increase from 3100MHz to 3200MHz must be absurd for me. I can do 3000MHz at 1.05V and 3100MHz at 1.1V.

Your SA and IO voltages need a bump for higher Mesh/Cache, try the values I provided and go from there. 3800Mhz Ram will most certainly need a bump in Uncore voltage aswell in range of +300

OH and for memory benchmark, I don't quite remember all, just tried pretty much anything google provided me with. Many benchmark incorporate memory tests along side their other tests but you'll get an idea. Never did I encounter low write performance like in Aida64 though, more like the opposite, very confusing.

[DooRules]