Optimizing RAM Performance for 3DMark Time Spy

Pro OC overclocker Alva “Lucky_n00b” Jonathan dives in to optimizing RAM performance for Futuremark’s latest DirectX12 benchmark Time Spy.

It’s been only a few weeks since Futuremark’s latest Benchmark, 3DMark Time Spy, was released to public. Now, overclockers and enthusiasts alike can start benching their system, and find out how their system compare to each other in that shiny new DX12 Benchmark – utilizing the modern APIs capability, like Multi-threaded Rendering, Explicit Multi-adapter, and Asynchronous Compute.

Normally, 3DMark are 3D-heavy, meaning that the performance effect of the GPU is much, much more apparent compared to the other system components, like CPU and RAM. This means an Core i7 6950X paired with GeForce GTX 950 will definitely give a much lower overall score compared a PC with Core i7-6700K with GeForce GTX 980 Ti. It doesn’t matter how much you push your CPU Scores by overclocking your CPU and RAM, it won’t give good scores if you have a weak GPU.

But there’s some scenarios where some components other than the GPU can give a small yet nice ‘boost’ to the overall scores. As a competitive overclocker who really cares about squeezing every bit of performance out of my system, I will explain how a simply optimizing RAM performance can give a little extra overall scores in the 3DMark Time Spy benchmark.

Ok, here we go!

Time Spy Optimizing RAM Performance Testbed

Time Spy Optimizing RAM Performance Testbed

 

The system that I’m using today consists of:

Here’s the normal 3DMark Time Spy scores:

Optimizing RAM Performance for 3DMark Time Spy Benchmark

3DMark Time Spy scores *click to enlarge*

Take notice that aside from the overall score, 3DMark Time Spy has 2 sub-scores:

  • Graphics Score
  • CPU Test Score

I won’t go into the detail as how Futuremark calculates 3DMark Time Spy scores, but their scoring method give good rewards if we boost whichever sub-scores that was slower. (You can read the nice technical guide from Futuremark if you’re into it – page 22)

For example, the score I put above have:

  • Graphics Score around 6200
  • CPU Test Score around 5384

According to the Futuremark Guide, in this specific scenario where the CPU Test Score is lower than the Graphics Score, we can boost the Overall Score higher if we increase the CPU test Score, by overclocking the CPU, (which I did, already at 4.5Ghz) and you’ve guessed it – we can improve our score by optimizing RAM performance.

For CPU test, 3DMark Time Spy used the Bullet Open Source Physics library, and aside for the CPU Performance, it can be affected by Memory Performance as well (yay, another memory benchmark!).

I’ve tried a couple of RAM timings and Frequency configs on my 2x4GB HyperX Fury DDR4-2666, from the standard DDR4-2133, up to DDR4-3600. The configs I use are listed here:

  • DDR4-2133, 15-15-15-36 2T, AUTO Subtiming
  • DDR4-2666, 15-17-17-36 2T, AUTO Subtiming
  • DDR4-3200, 15-17-17-36 2T, AUTO Subtiming
  • DDR4-3200, 12-16-16-28 1T, AUTO Subtiming
  • DDR4-3200, 12-16-16-28 1T, Optimized Subtiming
  • DDR4-3466, 13-18-18-28 1T, Optimized Subtiming
  • DDR4-3600, 13-18-18-28 1T, Optimized Subtiming

Here I’ll show you how optimizing RAM performance can affect both the CPU Score and the Overall Scores:

timespy_ramperformance3

 

In my tests, the effect of the RAM configuration definitely matter to the CPU Scores part(red bar), and it affects the Overall Score a little(blue bar). Notice that even with same RAM Frequency (DDR4-3200), optimizing both the main timings and the subtimings can give a little extra to the scores.

Here I don’t tweak the sub-timing as harsh as possible, mostly only tightening the tRRD, tWR, and tRTP.  Here are the detailed sub-timing for that particular run, as detected by ASRock Timing Configurator (also applied to the 3466 and 3600 configs).

Screenshot (173)

As addition, there’s two other setups (the DDR4-3466 and DDR4-3600) in which I intentionally gave looser timing, to show that keeping the timings tighten up can be pretty important as well, and you can see although the DDR4-3600 CL13-18-18-28 1T setup wins in CPU scores, the difference is not so much compared to the DDR4-3200CL12-16-16-28 1T config.

DSC08367s

So, there you have it. If you wanted to maximize your 3DMark Time Spy scores, aside from overclocking your GPU, make sure to pay attention to the CPU Test Scores as well – especially if your GPU Score is higher than the CPU Test Scores.

I haven’t tested the RAM effect on other platform, like Broadwell-E LGA2011v3 just yet, and I might do that next time if I have the platform ready to test. Hopefully that info helped you getting better scores, keep pushing it!

(Disclaimer: Alva Jonathan is part of the HWBOT Pro OC Program, a cooperative platform for the promotion of overclocking, its community, and its partners.)

330,984 total views, 17 views today

5 comments

  • Dragonmike

    Interesting. Have you checked whether the results in Time Spy are consistent with Aida64’s ram tests? Been using this to tweak my timings.

    • n00b

      When you’re tuning the main or secondary timings in RAM, the changes in AIDA Bandwidth test will be more apparent, and in memory-intensive test like Time Spy CPU Score (or 3DMark11 Physics Score), more AIDA bandwidth usually meant more CPU Scores.

      But when fine-tuning the tertiary timings (the tRDRD_sg, _dg, etc) and RTL, sometimes the changes are very small to be noticed by AIDA Bandwidth test (sometimes the ‘latency’ test shows it, sometimes it doesn’t), so it’s quite hard just to look from AIDA when doing this kind of tweaking.

  • Dragonmike

    Two more questions:
    1/ why did you use dual channel instead of quad? Would the gains just be the same?
    2/ what voltage did you give those sticks, and how did you cool them?

  • I haven’t always been looking at the CPU scores but now i have enough reason to run these tests. Thanks

Leave a Reply

Your email address will not be published. Required fields are marked *