Share your AIDA 64 cache and memory benchmark here

[mrcardio33x]

I think I'm done now.
VDD: 1.48v VDDQ/IO: 1.30v VSOC: 1.25v VDDP: 1.05v
VDDG CCD: 1.050v VDDG IOD: 0.950v - 2200 FCLK Stability.

 

[Sabotaged_Enigma]

Got something good
I've got a pair of 2933c13-capable Ramaxel 4 GB with Samsung E-die. Shame I didn't use them when overclocking 2600K. It's more like "Burning Bridge" at this frequency lol.

 

[fusseli]

Stock XMP:

 

[Paramount]

My result

 

[tugrul_SIMD]

gskill trident z5 neo ddr5 6000MHz CL32 1.35V expo enabled


just default expo + 2100 FCLK, not sure if its stable. bios always freezes no matter what even jedec (same as old ddr5 samsung)

 

[mrcardio33x]

A bit better - Only thing changed is used per core boost limit on my weaker cores (6,7,8) via (-) 100-200Mhz
VDD: 1.50v VDDQ/IO: 1.30v VDDG CCD: 1.050v VDDG IOD: 0.950v

 

[tugrul_SIMD]

One last optimization before tuning timings (except tREFI=50000, gdm off) of this trident z5 neo 6000-32-38-38-96:

If MSI is right, "highe efficiency" mode will give 12% more performance. Maybe I can try it today.

Edit: if you get BSOD right after windows login, its because of "power down" mode disabled with "memory context restore" enabled. So disabling both or enabling both solves BSOD.

Edit 2: fabric is 2133MHz

-----

Some loosened version of Buildzoid's timings harmed latency but gave 0.6GB/s reading & 3GB/s copying performance. Perhaps it was meant to be used with exact timing values? I just wanted to try a "safety" buffer zone on every timing (some values were already lower by default so I did not touch them).

Also "high efficiency" mode in bios did nothing. It had exact same timings as the default EXPO of the ram.

-----

So I reverted all the Buildzoid's timings back to default EXPO and tightened primary timings to 30-36-36 aaaaannnnddddd same happened, no performance improvement:

So, secondaries/tertiaries do not work, primaries do not work. There must be something missing. I don't know what it is... For now, I will play CS2 to test this CL30 thing stability.

Edit: with CL30-36-36, CS2 had 382 max FPS. Now I'll try with 20 millivolts extra.

With 1.37V dimm voltage (+20mV over default), CS2 had 392 max FPS and this:

imo it needs a bit more volts.

Edit: 1.38V did not improve aida64 much (only -0.2 nanoseconds) and only added 1-2 max FPS to CS2. Maybe this just made it stable. Now I'll try Buildzoid's timings on top of this.

Edit: 1.39V + 2133MHz FCLK (Buildzoid's timings did not work. I'm gonna learn some basic DDR5 timing rules later to not make mistakes):

Also counterstrike had 400 max FPS with this. +40mV should not be very frightening to use 7/24 I guess.

 

[mrcardio33x]

One last optimization before tuning timings (except tREFI=50000, gdm off) of this trident z5 neo 6000-32-38-38-96:

View attachment 326719
If MSI is right, "highe efficiency" mode will give 12% more performance. Maybe I can try it today.

Edit: if you get BSOD right after windows login, its because of "power down" mode disabled with "memory context restore" enabled. So disabling both or enabling both solves BSOD.

Edit 2: fabric is 2133MHz

-----

Some loosened version of Buildzoid's timings harmed latency but gave 0.6GB/s reading & 3GB/s copying performance. Perhaps it was meant to be used with exact timing values? I just wanted to try a "safety" buffer zone on every timing (some values were already lower by default so I did not touch them).

Also "high efficiency" mode in bios did nothing. It had exact same timings as the default EXPO of the ram.

View attachment 326749

-----

So I reverted all the Buildzoid's timings back to default EXPO and tightened primary timings to 30-36-36 aaaaannnnddddd same happened, no performance improvement:

View attachment 326756

So, secondaries/tertiaries do not work, primaries do not work. There must be something missing. I don't know what it is... For now, I will play CS2 to test this CL30 thing stability.

Edit: with CL30-36-36, CS2 had 382 max FPS. Now I'll try with 20 millivolts extra.

With 1.37V dimm voltage (+20mV over default), CS2 had 392 max FPS and this:

View attachment 326766
imo it needs a bit more volts.

Edit: 1.38V did not improve aida64 much (only -0.2 nanoseconds) and only added 1-2 max FPS to CS2. Maybe this just made it stable. Now I'll try Buildzoid's timings on top of this.

tRAS: 30
tRC: 36 - Yes you can run / be stable w/ lower than tRP+tRAS (ASUS has presets with tRC = tRP)
tRRDS: 4 if not 5,6.
tRRDL: 6 if not 7,8
tFAW: 16
tWTRS: 2 if not 3,4
tWTRL: 16
tRFC: 500 lower would likely be fine but won't really matter
tRDRDSCL: 4
tWRWRSCL: 4, 3 or 2 would likely work
tRTP: 12
tWRRD: 4, 3 or 2 might work

My prior R9 7900X with the same RAM kit, setup exactly the same as current:
Read: ~ 86000MB/s
Write: ~ 93000MB/s
Copy: ~ 84000MB/s
Latency: 57.2 - Current timings would likely be ~ 54-55ns. R7 7800X3D is roughly 4-5ns slower with same RAM/ timings.

You'll probably need to raise VDD to ~ 1.45v

 

[tugrul_SIMD]

You'll probably need to raise VDD to ~ 1.45v

mrcardio33x, did you mean this voltage for the settings you listed above, or was it for buildzoid's exact timings? I'm going to do research about how to set some timings not to increase latency accidentally, like this equals sum of that and that, etc.

Also you mean 1.45V for just VDD? Because when I type it, bios auto-sets 2 other voltages automatically.

Edit: is VDDQ/IO: 1.39v dangerous?

 

[mrcardio33x]

mrcardio33x, did you mean this voltage for the settings you listed above, or was it for buildzoid's exact timings? I'm going to do research about how to set some timings not to increase latency accidentally, like this equals sum of that and that, etc.

Also you mean 1.45V for just VDD? Because when I type it, bios auto-sets 2 other voltages automatically.

You'll need to enable the high voltage mode (don't know where it's located in your BIOS but should be there somewhere) Also, other voltages may or may not need to be adjusted, various MB to MB, RAM to RAM, etc.
Unfortunately just takes trail and error.. Just thought I'd give you some suggestions to try.

For my settings: VDD: 1.50v VDDQ/IO: 1.30v VDDP: 1.050v VSOC: 1.25v

At CL30 1.40-1.45v VDD is probably fine. But again these things vary MB to MB, etc.
Anyway, good luck you look to be heading in the right direction / not looking bad at all.

 

[tugrul_SIMD]

What is the difference between "high voltage" mode and just writing the voltage by hand?

 

[mrcardio33x]

What is the difference between "high voltage" mode and just writing the voltage by hand?

Many MB's won't allow entries higher than a certain value unless high voltage mode is enabled. 1.435v VDD seems to be the default highest allowed without enabling high voltage mode.
If high voltage mode (for VDD) changes other voltages you don't want changed just manually change those values to whatever you want to try.

 

[tugrul_SIMD]

"Tightest" version of "high efficiency" (the trick was to leave everything on auto except 3 primaries & tREFI) mode improved performance a bit (roughly equivalent to +25000 tREFI). But it did not touch some of timings good enough imo. What should I try in the red markers? Also yellow marker is different per dimm, one is 21 instead of 20. What to enter in this?

This is hynix-a die with 1.35V default, overvolted to 1.39V. Green parts are hand-tuned already (that tREFI allows the tightening work on high temperatures like 63Celcius at least until I plug a cooler onto ram sticks).

0.8 (aida64) nanoseconds latency improvement with tWR, tRC, tRAS.

 

[fusseli]

The plot thickens as stable tweaks are starting to settle in...

 

[Petar666]

 

[Gica]

View attachment 327082View attachment 327083

Bleah!

The real strength is here

 

[freeagent]

Ehh.. still a dinosaur no matter how many times I run it

 

[Det0x]

Fully stable 2:1 settings

Direct die 7950X @ large custom watercooling
2209mhz FCLK
8000MT/s @ 32-45-38-44-82

13k Karhu @ 338.94mb/s

http://imgur.com/a/s3Fk1Ip

And some bench 1:1 settings: (haven't uploaded to hwbot yet)

Direct die 7950X @ large custom watercooling
2202mhz FCLK
6725MT/s @ 26-36-32-30-62

 

[tugrul_SIMD]

What do you use for latency test? Fully random or sequential? Also how can you enter "1" to too many timings with just 1.2V dram?

 

[mrcardio33x]

Finally settled on a nice daily profile:

VDD: 1.435v VDDQ/IO: 1.30v VSOC: 1.20v

 

[tugrul_SIMD]

Does VDDIO help in stability about timings/subtimings or the connection like memory controller / DRAM controller? Is it ok if its same as VDD or 0.25V less than VDD or more than VDD?

--- 5200MHz DDR5 ---

Advantages against 6000MHz version on same tomahawk mobo:

- tREFI critical temperature is not reached instead of reached
- 40 seconds boot up without dram leraning instead of 240 seconds
- no file corruption
- 2200 MHz FCLK instead of 1800 MHz
- about 10% higher timespy cpu score

Disadvantages:

- 8GB/s less bandwidth
- 0.6 ns more latency

Performance changes during browsing.

 

[natr0n]

Got some really good B-die yesterday. I'm still testing. HappyNewYear

 

[Launcestonian]

View attachment 327883
Got some really good B-die yesterday. I'm still testing. HappyNewYear

There are no hard & fast answers about the right voltages to those rails, however general numbers can be used as a guide. Numerous sources around the web... silicon lottery plays a big part too.
Sometimes you can get away with 0.10v less between VDD & VDDQ depending on the sticks & ICs for example, you'll just have to experiment with your particular setup & see what works.

 

[fusseli]

I think I'm finally done for now, until Gigabyte drops a new bios and I have to start all over again like I just did

 

[Szabcsi]

Installing a new db server.