DDR4 Overclocking Guide

[RobertR1]

https://github.com/integralfx/MemTestHelper/blob/master/DDR4 OC Guide.md

One of the main areas where RAM overclocking helps is with minimum FPS. For VR users in Simracing, that's quite important to keep your minimum's up to avoid re-projection and stuttering. DDR4 tuning will do more for your minimum's then oc'ing your CPU in those instances.

I would strongly recommend reading through the entire guide and clicking on the hot links to learn the different concepts before diving in if you're new to this. *Do not let your motherboard set auto voltages when overclocking*

RAM tuning is like car tuning. It's not a set and forget. You need to play with combination of the settings and then test them for stability. It can take time so be patient. However, once dialed in, you can carry on normal operations and the system will be stable 24/7.

If you have questions about the guide or are stuck, feel free to post here. I personally don't have AMD experience but I'm sure others can chime in.

 

[RCHeliguy]

My Gigabyte Designaire MB had presets of the settings required for different OC DRAM by manufacurer. It immediately recognized mine and then changed to the appropriate settings immediately. Granted that is an Z390 Intel based MB.

It has been perfectly reliable.

 

[RobertR1]

that’s qvl and xmp. This is to go well beyond that especially on timings. Manufacturers set loose timings given a spec to get the highest compatibility across many boards and greater chance to work with most cpu imc’s.

 

[Vintagedriver01]

Thanks remembers me to ask an RAM question: as what would be the best choice for my new i7-9700kf build.
3200mhz cl16 trident from G.Skill and overclocked to 3333mhz or stay to
3333mhz cl16 Kingston at stock.
Gskill: https://tweakers.net/pricewatch/466589/g-punt-skill-trident-z-f4-3200c16d-16gtzb/specificaties/
Kingston: https://tweakers.net/pricewatch/551335/kingston-hyperx-predator-hx433c16pb3k2-16/specificaties/

 

[RobertR1]

The kingston is slightly better but neither are meant to overclock so get the ones that you like the looks of and fits your budget.

Kingston: DDR4-3333 CL16-18-18 @1.35V
GSkill: 3200MHz CL16-18-18-38 1.35V

 

[RCHeliguy]

I was just reading some testing that showed running dual channel memory vs. single channel memory had very little impact on games in terms of fps. I was a bit surprised about that. I would have thought it would be like RAID 0, but apparently not.

 

[dud]

Is that theory about raising minimum framerate supported by evidence such as measurements?

If the theory is correct then underclocking existing DDR4 should result in a noticeable drop in minimum framerate. So you could test it without going through the trouble of overclocking.

 

[Case_]

I was just reading some testing that showed running dual channel memory vs. single channel memory had very little impact on games in terms of fps. I was a bit surprised about that. I would have thought it would be like RAID 0, but apparently not.

 

[RCHeliguy]

Glad to see that since I have mine set up as 2 x Dual channel !

 

[dud]

Memory bandwidth and memory latency have very little impact on almost all real-world applications. Dual-, triple- and quad-channel RAM do act like RAID0 and drastically improve bandwidth. It is just that memory bandwidth doesn't have much of an impact on observable performance in what people typically do when they stare at their screens (not counting cracking passwords and things).

People can easily test that without the trouble and risk from overclocking by underclocking. You can underclock simply by setting a lower memory divider in the BIOS. That is quick and easily and is unlikely to threaten data integrity.

In an effort to be a little more productive: for the common sims, what is a widely accessible benchmark that displays minimum framerate? Just Assetto Corsa in benchmark mode? I have previously seem discrepancy in fps display between AC's benchmark display and other tools.

 

[RCHeliguy]

I was able to add additional super sampling when I bumped my MB's stock 2166 MHz configuration to 3200 MHz. I saw about a 10-15% improvement, but only because I was single core CPU limited.

I doubt outside of VR it would be a critical issue.

 

[RasmusP]

In an effort to be a little more productive: for the common sims, what is a widely accessible benchmark that displays minimum framerate? Just Assetto Corsa in benchmark mode?

AC is the only sim with benchmark mode. The minimum fps are laughable though as they start right at the beginning where everything is still loading so you might get hiccups of 1-3 fps, which will then be logged in for the min fps..

Easiest tool imo is msi afterburner + rtss, go into properties and "Benchmark", set the two buttons for start and stop and then run the ac benchmark.
Wait until your fps are up and running normally, hit start and stop it before it ends.

It will give you 0.1% low, 1% low and a not documented "minimum", which feels like some "5% low" since it's higher than the 1% low...
And average and maximum too!

 

[dud]

Yeah, that is what I remember. The AC benchmark runs in a reasonable manner, but the numbers are useless.

Do you remember offhand whether Python mods run during the benchmark? I have a little mod that collects all kinds of data, including frame timing. I just don't recall right now whether AC's benchmark drives on-screen mods and/or random python code that doesn't display to the screen.

 

[Case_]

That's generally the problem with minimum (and maximum) framerates - they can include absolutely useless values and you have to be mindful about what you're looking at. Like that Unigine benchmark we're running on S3 forums - the maximum framerate values fluctuate wildly and don't really tell you all that much, because they are generally logged during the scene transitions, not during the important parts. Similar with the minimum framerates.

 

[RobertR1]

Is that theory about raising minimum framerate supported by evidence such as measurements?

If the theory is correct then underclocking existing DDR4 should result in a noticeable drop in minimum framerate. So you could test it without going through the trouble of overclocking.

fpsVR is an overlay tool from the steam store that works in the HMD like an app in AC for example. I only drive in VR so thats' where the minimum jump helps and is easily noticeable especially since the current crop of sims are not multicore efficient. For flat screen gaming, especially with g-sync/freesync, hard to care. If my flatscreen game of BFV dips from 120fps to 111fps, it's fine. However, AC going from 90fps in VR to 80fps is noticeable and can add up to an unpleasant experience.

You can test this with an under clock but you might be moving other bottlenecks and the scaling might not be linear. Example, you can reasonably see a notable jump going from 2667 to 3200 mhz but from 3200mhz to 4000mhz it might not scale as well. Same goes for timings. CL14 at 3600mhz is a different ballgame than CL14 3200. Mem tuning is tedious and requires a fair bit of reading plus trial and error.

However, if someone still wants to try it, that guide is by the far the best reference material I've come across.

 

[dud]

Yeah, this is why I want to use my own code. Measuring the extreme top and bottom as such doesn't do any good. I want a proper log of every frame's timing and then do my own statistics on it. I'll see whether AC's benchmark drives Python mod code.

I am quite familiar with memory overclocking. I hold a second place world record, back from the DDR1 world. Geil One was one hell of a memory selection.

 

[RobertR1]

Update:

After getting my ram and cpu tuned, I now able to run AC at the following on my pimax 5k+:
- 72hz large fov locked at 72hz. Before it would dip at race starts and full pits scenarios
- 200% steam supersampling over 150% before
- PiTools at 1.0. Same.
- Same ingame settings at before (all max, 8xAA, 16x af, no post processing and no smoke. I hate the visuals of the smoke).

The locked fps was verified the by in-game renderstats app and content manager previous session performance feedback. During this testing I kept my nvidia driver version and pitools software at the same version as a control.

For me the tuning was well worth it. I could go beyond 200% but we're quickly at diminishing returns. For VR, the IQ is great and a notable jump over the previous 150% supersampling which also had FPS drops at race starts. Everything is as smooth as it can be. I did a 2hr+ continuous stint the other night in the Nring GP layout here on RD and absolutely no hitches at all.

Even rfactor2 is miles better but admittedly I haven't spent nearly as much time in that game to provide an honest comparison. Partially due to the nature of the game itself. However I was able to go from normal to large fov in rf2 and have no dips on mid ohio running the tatuus which is a nice change.

Current running config: https://valid.x86.fr/ucxyp7