How much is ram overclock potential influenced by CPU?

[mackerel]

A question for the more experienced at DDR4 overclocking, particularly with Skylake. How much influence does the individual CPU have on the overall stability when overclocking ram? This question first entered my mind when looking at the Asus QVL list for my mobo, where it says "*When running XMP at DDR4 3200 MHz or higher, the system’s stability depends on the CPU’s capabilities." Is that the case?

 

[ATMINSIDE]

It says that because of the IMC, not because of the CPU overclock.
The IMC is built in to the CPU.

 

[mackerel]

Understood. The question remains, is there that much variability in the IMC?

 

[ATMINSIDE]

Most can hit at least 3400-3600 with a decent set of RAM. The IMC will limit your max frequency at some point though.

 

[Woomack]

Nearly all Z170 motherboards have info in specs that 3200+ is never guaranteed. Most motherboards are actually tunned for full stability and won't run at much more than ~3600 while guaranteed max is still 3200.
So far I haven't seen IMC limiting max clock but motherboards. I can make 4250+ on 6600K and 6700K. On both memory is acting exactly the same. For sure there is the limit but for now bigger issue is to find good motherboard than good CPU.

 

[mackerel]

Thanks, looks like I don't need to worry any time soon. While looking for something like a beginners guide to extreme ram overclocking, I did come across Woomack's comment in another thread about dual rank having higher performance than single rank. Is there an easy way to tell which any set of ram is? I'm wondering if that could be contributing to ongoing performance differences I'm seeing.

 

[Woomack]

Dual rank is faster in case of DDR3, not really in DDR4. Just after premiere of Skylake I made a thread with comparison where dual rank DDR4 was slightly faster but results on newer series memory suggest that difference in performance depends more from used memory IC and SPD/XMP profile rather than single/dual rank memory design. I was testing that on 8GB single and dual rank modules and results were almost the same.
Right now the best DDR4 for competitive benchmarking is single rank.

 

[mackerel]

I was wondering if that could be a contributing factor in my testing in the other thread. Between 2 different G.skill kits on the same mobo, one is ~15% faster at the same clock and major timings than the other. Next to check out were the minor timings anyway.

 

[mackerel]

After much testing, which I touched upon in the other thread, I have some numbers to share.

I did testing with one system, an i7-6700k at 4.2 GHz, HT off, 4.1 GHz cache, G.Skill F4-3333C16-4GRRD at 3000 16-18-18-38. This is a 4x4GB kit. For each line, first value is for 4 modules fitted, 2nd value for 2 modules fitted (still in dual channel mode), and last one is first divided by 2nd to show relative performance.


Prime95 28.7 built in benchmark, 4 workers throughput in iterations/second at various FFT sizes. Run once each.
1024k 997.60 819.47 1.22
2048k 468.45 376.43 1.24
4096k 228.53 182.85 1.25
8192k 111.84 86.51 1.29

Here's the performance I was missing previously. 22% and rising with increasing workload. 1024k FFT would be 8MB data, x4 for each thread so this hits ram hard. Bigger FFT, more ram. It wasn't the motherboard. It wasn't the motherboard settings. It was the ram population. But is this effect seen in other benchmarks too? I tried a couple others as follows.


MaxxMEM2 1.99 (run 3 times, best results taken)
copy 33808 32634 1.04
read 26034 25583 1.02
write 30956 30956 1.00
latency 54.2 52.3 1.04

4% increase in copy, and 2% in read isn't much. Write unchanged. Latency increased though.


PassMark PerformanceTest 8 (ram tests only, run 3 times, best results taken)
Memory Database Operations 122.6 124.2 0.99
Memory Read Cached 31912 31922 1.00
Memory Read Uncached 20204 20244 1.00
Memory Write 16117 16397 0.98
Memory Latency 19.7 19 1.04
Memory Threaded 35007 34503 1.01

We see the same latency increase as with MaxxMem, but the rest doesn't change significantly.

So, any explanations for these observations? Is it the rank thing? Bare in mind I only care about prime finding performance, not any other benchmarks. It has long been known that as task sizes get bigger, the demands on ram performance increases. But I'm not aware of anyone trying to figure out how that works in practice. Real world testing sometimes even shows running 4 tasks will give lower throughput than running 3, although in most cases running 4 does give you more it isn't much more than 3 since it is so ram limited. From what I've seen so far, latency and processor cache size doesn't seem to play a significant role for these big units, although I haven't tested it in depth. Bandwidth seems to be king, but what is changing going from 2 modules fitted to 4?