My first serious mem OC attempt

As subject really.

Test system:
MSI Gaming Pro, bios 1.8
i5-6600k @ 3.5 GHz (for the purposes of this test, I'm only pushing ram and don't care what the CPU does) with Hyper 212 Evo cooler
ram: Corsair Vengeance CMK8GX4M2B3000C15 2x4GB - single rank, Thaiphoon burner isn't able to report what chips it uses
XMP timings are 3000 15-17-17-35...

First step, how high can I go? I turned on XMP and started turning up the clock. 3200 no problem. 3466 ok... 3600 post but not windows boot. Looked in settings, ram was set to 1.35v but reporting 1.344. I raised it to 1.36 and I got into Windows. As the speed went up, so did the timings.

So I'm at 3600 18-21-21-42... 1.36v set. A quick Prime95 bench (4096k FFT, 4 workers) gave a 3.3% throughput increase over 3000 XMP settings. Not great given 20% clock increase, but I am running a low CPU clock at the moment which might negate the benefit.

Can I do something with the timings? 18-18-18 appeared ok. When I tried to Prime95 stability test it, it failed quickly. Not good. I tried another minor voltage bump to 1.37v with no difference. Should I have just gone higher quicker? I backed off to 3466 at same timings, still Prime95 unstable.

I went next to 3200 16-18-18-38, and that seemed ok. Back to XMP voltage, now testing 16-16-16-36 timings and it looks ok. Do I try CL15 next?...

What more could I have tried when at 3600 for stability? I haven't touched things like VCCIO/SA at all yet. Or should I try tighter timings at lower speeds?

Try mine : 3600 15-15-1T 1.5v

So basically turn up the voltage and hope for the best? I'm doing a bit of stability testing on 3200 15-16-16 at moment which seems ok.

Ok, let's set 3600 again! Did you have to mess around with other settings?

Couldn't boot at 3600 1.4V at XMP/Auto timings, and I couldn't boot at 1.36 I had earlier either. Weird... made sure I didn't have left over settings too.

Going to take a break now, got it stress testing on 3200 14-16-16-36 while I'm away.

mackerel said:

So basically turn up the voltage and hope for the best? I'm doing a bit of stability testing on 3200 15-16-16 at moment which seems ok.

Ok, let's set 3600 again! Did you have to mess around with other settings?

Click to expand...

Nop, just annoyed a lot of peeps here with a lot of questions, then turned up the voltage and hoped for the best as soon as Woomack told me 1.5v is safe for 24/7 it was smooth sailing. Check out link below :

http://www.overclockers.com/forums/...us-VIII-Hero?p=7917584&viewfull=1#post7917584

Just tried 3600 1.5V, wasn't having any of it and wont post.

3200 C14 wasn't stable after all. It rebooted itself while I was away. Trying 3200 15-16-16-36 now. Hardly any increase over XMP 3000 15-17-17-35...

Just noticed, that the low profile version ? had a set of CML8GX3M2A1600C9 and never had any luck OC'ing them

3200 C15 unstable too. I think I'm going to call it a day here and revert back to standard XMP timings.

Still, it was interesting to go this far. There's probably more I could try but I have other things to do too

Oh, as for low profile or not... to me it is a standard size, not an oversize one, not a seriously shrunk one.

2x4GB and 2x8GB modules are on totally different IC and won't overclock the same. I'm 95% sure that this Corsair kit is based on Samsung IC ( but I can be wrong ).

4GB Samsung modules will overclock about:
- 3200 14-16-16 ~1.45V
- 3466 15-18-18 ~1.45-1.55V
not much more but some kits go higher at about ~1.5V
Hynix in most cases won't even boot above 3200/3333. So here you have the difference while testing memory.

2x8GB higher clocked single sided kits are on new Samsungs and are overclocking better at lower voltages. There is no way to compare any 2x4GB and 2x8GB kits as timings and required voltages are different.

Now you mention it, it sounds obvious I have a different kit... so in short, there isn't much to be gained there without piling on a ton of volts? It was a bit exciting to boot at 3600, even if I couldn't get back there or make it stable.

Is there a list of typically (DDR4) ram chips and their typical OC characteristics somewhere? I thought I saw something like that before but don't seem to be able to find it now.

I was a bit surprised to be shown a Crucial DDR4 4GB module that is claimed to be dual rank. Only comes in 2666C16 so not fast, but I've got a set on order to try it. Do Crucial only use Micron chips or do they use others also?

Crucial uses Microns as they're Micron brand. Well, in the past they used different IC in single series but at least in DDR4 all kits are based on 3-4 Micron D9 series. I made 3 reviews based on Crucial kits. All had Micron D9 inside and on all were different results.
Where did you find info that 4GB module is dual rank ?

I know Crucial are the consumer brand for Micron, but it isn't unknown for manufacturers in general to go elsewhere if there is a niche to be filled or better value option available.

http://uk.crucial.com/gbr/en/blt2c4g4d26afta

There is a risk the web page is in error, but regardless I've ordered from Amazon on free delivery so hopefully will get it later this week to check out.

A side question, I keep debating if I should buy AIDA64 mainly for the memory system benchmarks, although the testing charts may come in useful also. I've been using MaxxMem2 as a synthetic measure. How do their measures compare? I've used up the AIDA trial on my systems previously so can't go back to compare.

AIDA64 is multi-threaded while MaxxMem is only single-threaded. There are no modern single threaded OS anymore as everything is using load balance. AIDA64 is simply better but MaxxMem isn't bad if you compare memory kits.
If you wish other benchmarks for memory tests then you can use Geekbench ( full system test but has memory tests separated ) but free version is 32bit only or simply use "winsat mem" in command prompt and Windows will show you memory bandwidth.

Mackerel, Corsair is not interested in programing DRAM manufacturer ID within SPD bytes. This is the reason for which Thaiphoon Burner can't determine it. Corsair's SPD is always incomplete.

AIDA64 purchased. So guess what I'll be doing shortly?

In other news, remember I said earlier I was disappointed that going from 3000 to 3600 only showed 3.3% improvement? Well, I just went the other way and tested at 2400 12-14-14-35. Going from that to the XMP 3000 settings, I saw a 20.6% improvement for a 25% ram clock increase. It suggests I got most of the bandwidth I really needed for this CPU clock by doing that. Anything after that is hitting diminishing returns.

Side note: I tried to put this through my previous model of predicted ram performance. I long suspected it overestimated how much ram bandwidth you needed relative to CPU demand and sure enough here it predicted bigger scaling than seen based on single rank. Actually the dual rank results were closer. My rank tests were only done at a single ram speed and I took that at face value, which should be incorrect as I needed to factor in how that affects the overall balance, not just the relative result at that speed. I'll need to go away and think of a more suitable test, such as ensuring I'm still well in the ram limited region when varying rank.

Ramires said:

Mackerel, Corsair is not interested in programing DRAM manufacturer ID within SPD bytes. This is the reason for which Thaiphoon Burner can't determine it. Corsair's SPD is always incomplete.

Click to expand...

The same Mushkin and Patriot in latest series.

The so called dual rank ram has arrived, but I wont be in a position to test it until much later. In a visual inspection I can only see chips on one side of each module.

I'm going to guess based on the kit having just eight chips on one side... That the results of running a Prime v28.7 benchmark for the new kit will be similar to that of running with just 2x4GB from your G.Skill 4x4GB kit?

BTW... With my Crucial Ballistic Sport 2400C16 16GB (2x8GB) kit... The max speed seemed to be ~3000/3066 (~1.35v)... When trying for 3200(+) I was seeing no boot/failed posts.

Single rank. Haven't benched it yet. Only just plugged it in and tried to run Thaiphoon Burner, which took a while to get working as MSI Command Center was interfering with it until I uninstalled that. Without OC, I have to wonder if running 2666 dual rank (4x4GB) is faster than 3000 single rank (2x4GB). According to my understanding up to now, it should be, but I will need to prove it again. Also I'll be mixing ram spec if going 4 modules so dunno if that might cause problems. This will be a supertest I'll likely do on the weekend as I want to go in more detail with AIDA64 now I have it.

Here is a Thaiphoon Burner/AIDA64 for my Crucial Ballistix Sport 2400C16 2x8GB kit...

6700K 4.2GHz 3000C14 (1.3v DRAM; 1.1v VCCIO; 1.17v VCCSA):