About G.Skill Trident X 2400

[DonkeyOatie]

A parent donated a set of 2 x4 and 2 x 8 G,Skill Trident X for the Science Fair computer. In order to ensure that my students projects' will discover something interesting, I have been running a lot of tests on the memory in various configurations. I have observed that 2 x 4 is somewhat faster than 2 x 8, using my G3258 CPU on an ASRock Z97M OC Formula board, at every memory speed from, 1333 Mhz to 2400 Mhz, (using other identical settings) and at processor speeds from 3.2 Ghz to 4.2 Ghz, with and without a discrete GPU.

Anyone got a clue why? (are 8Gb stick single rank, and 4Gb dual rank?)

Thanks.

 

[clonazepam]

The processor is a limiting factor. The tests themselves may be inadequate for benchmarking RAM (limited by another factor). Lastly, it's possible the 8GB sticks are run with slower tertiary settings to maintain compatibility.

Sounds like a fun time with the Science Fair and great to hear about such involved parents. That's pretty cool.

 

[DonkeyOatie]

Can you explain why or how the CPU would cause 2 x 8 to run slower than 2 x 4? I'm not actually benchmarking the RAM, I'm looking at the effect of of different RAM configurations 1 x 4 ,2 x4, 1 x 8, 2 x 8, 1333, 1600, 2400 Mhz, and 3.2 Ghz and 4.2 Ghz in all possible combinations on a series of benchmarks.

I'm using the same timings for all tests. Perhaps the 2 x 8 needs different timings to run faster?

Here's the sort of data I'm collecting.

These runs show the effect of CPU speed on the benchmark performance.

 

[Tradesman1]

The 2x4GB isn't as much stress on the MC as running 2x8GB. If possible and can 4x4GB vs the 2x8GB you may well find that the 4x4GB are then slightly lower performance wise, as while both are 16GB the 4x4GB is more stress than the lesser sticks in the 2x8GB. Can often balance performance out with small amounts of additional voltage to the DRAM and/or MC (memory controller)

 

[DonkeyOatie]

Thanks Tradesman1. I don't have 4 x 4Gb to try.

Here's my raw data.

and with GPU

 

[Tradesman1]

Yeah, it can come up with some interesting figures. I did a ton of testing with about 8 32GB sets of DRAM for an article and it seemed to take forever, ran them in 1x8GB 2x8GB and 4x8GB

http://www.tomshardware.co.uk/dram-benchmark-fluctuations,review-33154.html

 

[DonkeyOatie]

With the overclocking, burn-in testing and RAM testing it takes about two days per data set.

So you still think that the limit is the CPU MC, and it cannot handle 2 x 8 as well as 2 x 4?

 

[gskill support]

Not exactly, you have to consider 8GB is twice the capacity as 4GB. It's natural latency is increased and performance is slightly decreased due to a greater capacity. The CPU also receives more stress/load with a higher capacity so there are many factors.

Take our world record overclocking for example, the record is DDR4-4431 with a single 4GB stick, DDR4-3872 with 16GB.

http://www.gskill.com/en/press/view/g-skill-memory-broke-18-overclocking-records-since-computex-week

As you can see, the difference is significant. This is not to say we should all run a single stick in our computers because with dual and quad channel memory architectures built with the motherboard, bandwidth is multiplied by having more than 1 module. The increase in bandwidth is a greater useable advantage than high DRAM frequency.

 

[Tradesman1]

With the overclocking, burn-in testing and RAM testing it takes about two days per data set.

So you still think that the limit is the CPU MC, and it cannot handle 2 x 8 as well as 2 x 4?

_________________

GSkill went into a little more depth, but same basic thing a little additional stress of the CPU which the MC is a part of, and as I mention adding a little additional voltage to the 'larger', whether it be more DRAM or more sticks can help level things out, can even at times increase performance of the 'smaller' - or say the 2x4GB (in the it vs 2x8GB, where the 2x4GB did better)...There's a lot you can play with

 

[DonkeyOatie]

Thanks for the info. Eventually, we will be using an i5 4690 and an i7 4790 in the same series of tests, so we will see what happens.

I am aware that I can wring out more performance by adjusting voltages and perhaps by using different timings for the 2 x 4 and the 2 x 8. However, that is not the purpose of this project (but may be the focus of another project). This project is about identifying 'efficient' configurations, or determining the performance of specific configurations to assist with buying decisions.

I guess it's a sign that our methodology is good that we are detecting such small effects.

I hope the data we are collecting is useful and meaningful. There's a limit to what Middle School students will do.

 

[Tradesman1]

There's a limit for big folks also I've got some sets for testing and reviewing, OCing is going to be part of it, but am thinking of setting a limit on the time to be spent in OCing, of prob 2 hours at both 16GB and with the full 32GB load

 

[DonkeyOatie]

So far the information on performance, looking at overclocking, memory configurations, iGPU overclocking, and the various benchmarks has been very interesting. In particular, the Caselab Euler 3D fluid dynamics testing, with a spectrum of 1 - 8 threads, with no graphics use, shows small variances which gives an insight into how the CPU handles multiple threads. It will be interesting to see how the i5 and i7 test.

I'm thinking that I can call the system stable after 8hrs of Prime95, except at the highest overclocks. If it will run stable at 4.2Ghz, I se little point in heavy testing at 3.5 Ghz.

 

[Tradesman1]

Sounds like it's good