Upgrading a 2x8 to a 4x8 for G.SKILL 3200 TridentZ -- any last minute tips?

[BonzaiDuck]

I have been advised -- and I would advise myself -- to simply replace a 2x8=16GB kit with a 2x16=32GB kit if doubling the RAM was an objective.

I had also been in touch with G.SKILL e-mail tech-support, asking about the comparative possibilities. Their pronouncement began with a disclaimer that they couldn't offer "warranty replacement" or tech support for use of two identical F4-3200C14-8GTZ to augment the same. That is, if the dual 8GB kit ran fine by itself, RMA would be pointless.

When I mentioned down-clocking the stock "OC" 3200 speed, they responded with some positive enthusiasm. But that doesn't mean they won't run flawlessly at 3200, although I'd be perfectly happy to drop that clock to 3100 or 3000. I have even seen more ambitious folk touting the "over-clock-ability" of the same kits if run in a 16GB configuration.

I already arranged the RMA replacement of the second and new 16GB kit for a pair of F4-3200C14-16GTZ modules, resulting in an increase to my original credit-card charge. I just couldn't resist penny-pinching on my RAM upgrade, and I just wanted to "see." I've got 14 days to do that.

Others have mentioned that the use of four slots will stress the IMC of the processor a bit more, and I might have to bump up the VCORE a tad. Again -- I'll just have to see . . . .

Anyone have an idea how reducing RAM speed by 100 or 200Mhz will reduce the burn on the IMC?

And anyone have other insights? If this doesn't work out, as I said -- I can go forward with the RMA replacement and give up the idea of saving $180.

 

[BonzaiDuck]

Well . . . it's been maybe an hour or so . . . 21 people have viewed the post.

First, I'm reasonably impressed with HCI Memtest 64 Pro v 5.1.

I put the second pair of modules in the slots, reconnected and powered up to BIOS . . . it sees all 32,768 MB . . . CPU clocks are stock, but I kept the XMP 3200 14-14-14-34 setting. Nine instances of HCI Memtest are running -- spawning them with this version takes a single mouse-click. The new version boasts reducing test time by a factor of 2 or 3.

So . . . proof of the pudding may be exclusive to twisting the CPU clock back up to the 47 multiplier . . . Some of the instances have hit 10% coverage so far in just a few minutes.

We're gonna find out. I know I'd posted in the forums before looking for firsthand experience, and I agree that the advice was good. But for firsthand, nobody said much of anything . . .

 

[BonzaiDuck]

Pushing past 130% coverage for all ~3,000 MB instances of HCI Memtest, it looks as though this experiment is a success.

I can run the two kits of TridentZ at stock "OC" spec of 3200 and 14-14-14-34.

The rest of it will be a testing project to see if I can get back to 4.7Ghz for the CPU overclock at those settings, how much additional VCCIO and VCORE I'll need to make it stick, or whether dropping back to DDR4-3100 or 3000 will be sufficient to make it all work flawlessly without fiddling with the CPU and its IMC voltages.

I'm surprised I agonized so much over these choices or even the objective of increasing RAM. At least I postponed the expense for several months before today . . . I'll post my review at the Egg . . .

 

[Carfax83]

Skylake has a very robust memory controller, so it doesn't surprise me it was a success. If I hadn't been at work earlier, I would have told you that more DIMMs generally means a slight degrading in performance and stability depending on how aggressive the speed and timings are. But this HIGHLY depends on how robust the memory controller is, as well as the quality of the ICs. The Trident Z DIMMs use the Samsung B die ICs which is basically top of the line. I have the same RAM that you do, but in quad channel config and I run them at 3400 CL 15-15-15-34 CR1 timings at stock voltage.

So yeah, you lucked out. If you had crappy memory, I doubt it would have been successful.

 

[BonzaiDuck]

Skylake has a very robust memory controller, so it doesn't surprise me it was a success. If I hadn't been at work earlier, I would have told you that more DIMMs generally means a slight degrading in performance and stability depending on how aggressive the speed and timings are. But this HIGHLY depends on how robust the memory controller is, as well as the quality of the ICs. The Trident Z DIMMs use the Samsung B die ICs which is basically top of the line. I have the same RAM that you do, but in quad channel config and I run them at 3400 CL 15-15-15-34 CR1 timings at stock voltage.

So yeah, you lucked out. If you had crappy memory, I doubt it would have been successful.

Thanks for "firsthand observations." I wasn't planning to clock these higher: they are the TridentZ's of this speed with the tightest timings. But it's a good thing to know. Definitely.

UPDATE: I'll have to run these tests again, if only to satisfy myself.

I had a profile in my BIOS for "stock settings." The XMP profile was set appropriately; most everything was set to "Auto." Now . . . I'm 70 years old, once had 20-20 vision, and now use two sets of eye-glasses: a progressive lens pair for doing things in general, and a second prescription for use exclusively in reading or using my computers. I'm pretty sure I was wearing the former pair of glasses this morning.

When I went into BIOS to reset the profile to "4600 Mhz," I noticed that the MONITORED RAM voltage showed (what I thought was) 1.544V!!!! I also know that the fixed setting of 1.35V would typically show 1.344V in the monitor, and a "3" might be mistaken as a "5" with my early-morning vision assisted by the progressive lenses.

OCCT is running fine right now and clicking past 2 hours for the "CPU:OCCT" test, but that test only uses 1792MB of RAM. I could either proceed with my tests of the CPU overclock using the full memory under something like LinX or an incarnation of Prime95 and see if I run into trouble. Or I could back up and rerun HCI Memtest for 200% coverage to see that it's all Kosher.

I hate stupid mistakes, and worse -- I hate it when I'm not completely sure either way that I made one! The prevailing wisdom about DDR4 suggests that the upper limit of a "safe range" for vDIMM is 1.5V for 24/7 operation.

And maybe I'm just imagining all this, but that's the result of carelessness.

 

[BonzaiDuck]

Interesting. For months I solicited insight into this matter of separate dual-stick kits of the same model. I mentioned the cautions earlier. G.SKILL tech support was reserved in their remarks. Carfax83 explained more than I'd known about these kits.

I reset the system to my 4.6Ghz OC settings, and ran INtelBurnTest and then LinX for hours. Tight grouping of GLOPS results -- little variation. Didn't need to adjust anything for the extra memory kit. For DDR3 in my Z68 boards, I actually had to adjust VCORE.

I should've added this second kit months ago . . .

 

[BonzaiDuck]

Just for the record, I opened a thread on "CPUs and OC'ing" about the behavior of stress-tests.

I've come to a new realization that RealBench is not just a "limpy-limpo" test: it can uncover errors much earlier than LinX and without the "burn" -- and earlier than OCCT:CPU. I can get these modules (4x8) to run with the CPU at 4.7Ghz, but I either have to fix VCCIO at a higher level, raise the fixed VDIMM above 1.35V, or raise VCORE for the full spec ("OC") speed of 3200.

How much is 100Mhz in the CPU clock worth in tweaks and trouble? With 32GB of RAM, there's also an extra benefit over 16. Everything is a trade-off with something else.

AND ANOTHER: It's a bad habit that I post a lot when I'm doing something like this. About all this needed to run full-out at spec DDR4-3200 and 4.7 is a bump up in VCCIO -- somewhere between 1.16 and 1.20V. Nothing more than that.