[munternet]

As the title suggests setting rising and falling data slopes seems important for overclocking memory if you want to push it a little

If this seems familiar it's probably because I copied and pasted it from my last thread on skew control and changed a few words


I spent an afternoon testing different settings to see just how much difference it makes and it seems it won't even boot with many of them

I can see now that it would be easy to think you were unlucky in the silicon lottery or your IMC isn't any good just because your slopes aren't set correctly

The test I'm using for stability is GSAT. I was using TM5 also but it seems it isn't needed really and just makes the process take longer
There is a GSAT download link and some instructions in my sig if you want to try this for yourself


So my first step was to find a setting I knew to be not quite stable
You might be able to lower VCCSA a little if you aren't getting enough errors. Lowering VCCIO will probably cause freezes

5.2/49 1.32v LLC7
4600-16-17-17-36
1.58vdimm (water cooled)
1.43vccio
1.45vccsa

Next I divided the settings into 4 pairs and set them one pair at a time

Data Rising - Data Falling
CMD Rising - CMD Falling
Ctrl Rising - Ctrl Falling
Clk Rising - Clk Falling


After finding a number that worked with both the rising and falling of the pair I set about finding out which way the pair liked to be offset best. Maybe rising liked to be 1 or 2 values higher or 1 or 2 values lower than falling. Maybe 3.
Once I found the desired offset I could raise or lower the data rising and data falling together as a pair to where it errored least
After doing all 4 sets you can go through them again and see if it helps, but in my case I found the best values first time through


Here are the results

These results laid out in this manor make it pretty clear what my optimum options are for this hardware


Now with skews, slopes and Vref set it was very easy to tighten things up and boost the Aida scores

When I first started setting data slopes I was running 4500c16 daily and 4600c16 tested well some days but not others but now that I have completed the testing and setting it's possible for me to run 4600c16 ram overclock with 2x16GB sticks with similar voltages as my daily

On a side note I think other combinations such as 4600-17-18-18-38 would probably train now and run with lower voltages without much of a performance hit as @PhoenixMDA does

Boots are consistent and rock solid now where-as before I would get the occasional glitch or bad training etc but now it's as stable as when I'm running 4000MHz
My hardware is in my sig for reference

This worked for me on the Apex XII after setting ODT Skews (linked in my sig) and Vref which seems to be just a matter of finding the right value and locking it in (may need adjusting if you make major changes)

If you have some information that may be useful to others with different boards feel free to post it here

Open Office is a free download if anyone would like to try this at home and add your results to this thread for future reference
Be sure to list your parts in your sig or in your post, but it's well worth doing a rig builder in your sig anyway

I look forward to seeing your results

 

[YaqY]

As the title suggests setting rising and falling data slopes seems important for overclocking memory if you want to push it a little

If this seems familiar it's probably because I copied and pasted it from my last thread on skew control and changed a few words


I spent an afternoon testing different settings to see just how much difference it makes and it seems it won't even boot with many of them

I can see now that it would be easy to think you were unlucky in the silicon lottery or your IMC isn't any good just because your slopes aren't set correctly

The test I'm using for stability is GSAT. I was using TM5 also but it seems it isn't needed really and just makes the process take longer
There is a GSAT download link and some instructions in my sig if you want to try this for yourself


So my first step was to find a setting I knew to be not quite stable
You might be able to lower VCCSA a little if you aren't getting enough errors. Lowering VCCIO will probably cause freezes

5.2/49 1.32v LLC7
4600-16-17-17-36
1.58vdimm (water cooled)
1.43vccio
1.45vccsa

Next I divided the settings into 4 pairs and set them one pair at a time

Data Rising - Data Falling
CMD Rising - CMD Falling
Ctrl Rising - Ctrl Falling
Clk Rising - Clk Falling


After finding a number that worked with both the rising and falling of the pair I set about finding out which way the pair liked to be offset best. Maybe rising liked to be 1 or 2 values higher or 1 or 2 values lower than falling. Maybe 3.
Once I found the desired offset I could raise or lower the data rising and data falling together as a pair to where it errored least
After doing all 4 sets you can go through them again and see if it helps, but in my case I found the best values first time through


Here are the results

View attachment 2479878


These results laid out in this manor make it pretty clear what my optimum options are for this hardware





When I first started setting data slopes I was running 4500c16 daily and 4600c16 tested well some days but not others but now that I have completed the testing and setting it's possible for me to run 4600c16 ram overclock with 2x16GB sticks with similar voltages as my daily

On a side note I think other combinations such as 4600-17-18-18-38 would probably train now and run with lower voltages without much of a performance hit as @PhoenixMDA does

Boots are consistent and rock solid now where-as before I would get the occasional glitch or bad training etc but now it's as stable as when I'm running 4000MHz
My hardware is in my sig for reference

This worked for me on the Apex XII after setting ODT Skews and Vref
If you have some information that may be useful to others with different boards feel free to post it here

Open Office is a free download if anyone would like to try this at home and add your results to this thread for future reference
Be sure to list your parts in your sig or in your post, but it's well worth doing a rig builder in your sig anyway

I look forward to seeing your results

So my understanding is that with slope tuning that you have happened to get 4600C16 fully stable which was not possible before? Did you start this testing with tight subtimings/tertiaries or with loose timings?

 

[munternet]

So my understanding is that with slope tuning that you have happened to get 4600C16 fully stable which was not possible before? Did you start this testing with tight subtimings/tertiaries or with loose timings?

Prior to adjusting these settings I thought I was stable at 4600c16 but then for no apparent reason the training changed and there were errors. After setting slopes I was stable and able to tighten everything up substantially more than before without throwing insane amounts of voltage at it. I don't even get any bad boots now
Find an overclock with tight timings and tertiaries you think are correct and tighten RTLs until you see errors in GSAT. You may need to drop VCCSA a little to get the errors required to identify the sweet spots

 

[PhoenixMDA]

@munternet
Two things CL16 is much easier as CL17 for IMC.Its not said that you can get CL17 stable.
With 2 other 10900k (bad ones) for me it wasnt possible to get 4600CL17 really stable with my 2x16GB.
But 4500/4600CL16 was possible there.
You must test it.
Every Setting needs his own Slope's, for CL17 you will need probably other slopes.

The other thing is, every boot is random, that mean that the other slopes on auto can be good or not so good, so you cant be 100% sure that your best slope is really the best one.
Thats the reason why it's so difficult to find really the best.
If i have find my slopes, i begin in the front and testing again, with the fixed values, to find perhaps a better one.

See it so if the Board has problems to fixe good values the signal quality is on the limit, if you fixe the best ones you have a good chance to get it stable.
I testing with some boots GSat stability, if i see it can run good, then i search the slope's.

 

[8051]

Those VCCSA's and VCCIO's seem really high. Are those only safe/valid for Comet Lake CPU's?

 

[munternet]

Those VCCSA's and VCCIO's seem really high. Are those only safe/valid for Comet Lake CPU's?

Well within spec for this gen apparently

 

[moonbogg]

I was waiting for people to say "IO/SA voltage should be 1.0v or less under any and all circumstances". I didn't get what I was waiting for, but it came close enough. I love the SA/IO voltage hysteria. It's hilarious. 9 times out of 10 people are just bragging or trolling when they say things like that, such as "my 10700K is 5.3ghz with 1.2 vcore and you should easily get that too" or "OMG get that SA voltage below 1.1 MAXIMUM. My RAM is 5000mhz and I use 1.1 SA and you should need even less than me with your cheapass DDR 4000 kit".
As a matter of fact, I recommend just turning SA voltage off. Unless your motherboard is a piece of crap, you should see the setting under dolt volts.

 

[PhoenixMDA]

Things like Slope's and Bitlines are not for casual user and not for childish people.
I think the most people dont understand for what that is and it gives also more setting with influence.
The most only known voltages and subs...

I think thats enough information, more is not necessary, the rest the people have to work for themselves

People with really technical technical interests can all find in the associated patents.

 

[Nizzen]

Those VCCSA's and VCCIO's seem really high. Are those only safe/valid for Comet Lake CPU's?

Under 1.5vccsa is low with Apex and really tweaked b-die memory. 4600c16 and faster
If you want 73-74 GB/s read and sub 35ns memorylatency, you pretty much need 1.45v vccsa, unless you are running ln2

 

[YaqY]

Have been messing with slopes for a while now and cant seem to get any consistency, i like to check values 2-3 times for each combination of data rising for example, one pass it might do 600 seconds before error and the next boot it will do 10 seconds, don't understand the high inconsistency and a proper test methodology to this.