[QUOTE=chalsall;453154]Please forgive me for this, but have you run further experiments?

For example, are you absolutely _sure_ that that one module is at fault? Perhaps it's the motherboard or the controller. Or, perhaps, it simply isn't seated properly. Have you tried switching them around and see what happens?

My apologies if I am telling you how to chew gum. But I have found that crying wolf is not wise if a wolf is not present.[/QUOTE]
I appreciate shared experience, which such questions represent.

What I can say is that I started getting BSODs for Memory_Management. This led to web searches, and running Memtest with both sticks in, which caused an instant flood of errors. I tried first one DIMM and then the other in the first slot of the two I use: the second and fourth, as advised in the manual. One could run Memtest through a full cycle with no problem. The other delivered many errors instantly.

I didn't really see the point in going further, since one seemed pretty surely to be the problem.

[QUOTE=kladner;453166]I didn't really see the point in going further, since one seemed pretty surely to be the problem.[/QUOTE]

Agreed. It seems you have run the appropriate tests appropriately.

Sorry for questioning you. But far too often people (including myself) make conclusions without fully exploring the possibility space. You clearly have.

Namastey.

I truly appreciate the thought and effort. You never know if something might have been overlooked. Doesn't hurt to check.

EDIT: I do all the driving in this family. I like to, and Dan hates to drive. However, I work strongly to shut down any irritation I might feel if I have already seen something that Dan points out. I encourage him to say something on the premise, "Four eyes are better than two." Attention can wander on long drives. I much prefer warnings that might be unneeded, to having a crash.
This mindset carries over to other areas. It comes of trying to reduce irritability in general.

[QUOTE=kladner;451113]I exchanged the XMP 2400 parts for XMP 3000. Strangely, running a single, four thread worker is slightly (0.10 to 0.20 ms) slower, now. Would the increased speed show up if I were running more than one worker?

EDIT: I did notice last night that the current RAM shows Single ranks in CPUZ>SPD. I don't know what the previous RAM indicated. I am supposing that it would be the same as these are from the same Corsair Vengeance LPX product line.[/QUOTE]
Short update: it finally hit me that all I had to do was look at my order history. The original Kingston 2133 MHz parts are Dual Ranked. It seems that the Corsair are mostly Single. I am now trying to find something that Microcenter carries, or will order for me. There are Kingston 2400 MHz parts that are dual, though not listed with MC. Certain Crucial parts give the Rank information, and some of those are Dual. Micron makes some, but hardly anyone carries them.

Here is a list I compiled of published Dual rank specs. Double asterisk is Dual.
Kingston HyperX Fury:

HX424C15FBK2/16 **
HX421C14FBK2/16 **
HX424C15FB/16 **
HX424C15FB2K2/16 *
HX424C15FBK2/16 **
HX421C14FB2K2/1 *
HX426C15FBK2/16 **
HX424C15FB/8 **
HX424C15FB/16 **
HX424C15FB2/8 *
HX424C15FB/8 **
HX421C14FB/16 **
HX426C15FBK4/32 **
HX421C14FBK2/32 **
HX426C15FBK4/32 **
HX424C15FBK4/32 **
HX421C14FBK4/32 **
HX424C15FB/16 **
HX421C14FB/16 **
HX426C15FBK2/16 **
HX421C14FBK2/16 **
HX426C15FB/8 **
HX424C15FB/8 **
HX421C14FB/8 **


CT2K8G4DFD824A ** Crucial
CT2K16G4DFD8213 ** Crucial
CT2K16G4DFD824A ** Crucial

HX424C15FBK2/8 * Kingston
KVR21N15D8K2/32 ** Kingston
HX424C15FBK2/8 *
KVR21N15D8K2/32 **
KVR24N17D8/16 **
KVR21N15D8/16 **
KVR24N17S8/8 *

More Dual Rank info
 

Well, Microcenter does not have any of the known dual model numbers posted above. Worse, they don't order stuff that's not already stocked. The only ordering they do is for a listed item that is out of stock.

That said, I am back to ordering from NewEgg, or direct from either Crucial or Kingston. Kingston seems to be the fastest choice available with specific information on Rank. Basically, the Crucial (brand) DIMMs top out at 2400 MHz. Kingston has the same sized parts going to 2666. That, with the dual-rank boost should be impressive.

From earlier in the thread:
[QUOTE]Originally Posted by [B]kladner[/B] [URL="http://www.mersenneforum.org/showthread.php?p=451217#post451217"][IMG]http://www.mersenneforum.org/images/buttons/viewpost.gif[/IMG][/URL]
[I]One factor which might be in play here is that the original XMP-2400 RAM was CAS 14, while the XMP-3000 is CAS 15. The difference in ms/it is 2.628 versus ~2.877.[/I]
[/QUOTE]
I now believe that the real reason is that the very first RAM I got, Kingston DDR4-2133, XMP2400, was dual ranked. I have to surmise that the Corsair modules I replaced them with were single rank.

Microcenter only carries a few Kingston items. I picked that model for low latency. They did not even have it as a kit. I got two individual DIMMs. This is not an issue with either NE or the manufacturer-direct vendors. The kits do exist.

Some links follow. If they end in PDF they are spec sheets for relevant parts.
[url]http://www.kingston.com/dataSheets/HX426C15FBK4_32.pdf[/url]
[url]http://www.kingston.com/dataSheets/HX426C15FBK2_16.pdf[/url]
[url]http://www.kingston.com/dataSheets/HX424C15FB_8.pdf[/url]

This is the part I originally bought 2 of:
[url]http://www.kingston.com/dataSheets/HX421C14FB_8.pdf[/url].

Finally, some screen shots of Compare tables from various sources.

The conclusion I offer is that the HyperX Fury parts are a better choice. The Crucial DIMMs are ultra-retro bare chips on green circuit boards. The Kingstons have nice heat spreaders. :smile:

It ended up being NewEgg. Cheaper and faster than OEM.
HyperX Fury 16GB (2 x 8GB) DDR4 2666 Model# HX426C15FBK2/16-SLV

This is listed twice on NE, with ~$20 difference. I hope my assumption that Sleeved is the equivalent of a bare HDD is correct. That is, same product, less packaging.

I hope to see them tomorrow. I paid for faster shipping because I am coming up on the 30 day limit for store return of the pair with a failure.

If my previous performance comparison of DDR4 2400 dual rank to DDR4 3000 single rank DIMMs is real, then I have some hope that P95 at 2666 MHz will perform on par with 3200 MHz on single rank RAM. :w00t:
I wonder if this carries over to P-1.

Woohoo! Got the RAM, but delivered at work. Have to be patient until I get home ~3:30 PM :whee:

EDIT: Hold the presses! I just realized that the first two examples are of 39.9M exponents. The most recent is 42.5M. If I am not mistaken, this tips the balance even more toward the HyperX 2666 RAM.
See below:

This particular HyperX Fury 2666 is interesting, in that JDEC and XMP specs are identical.

"Sleeved" turned out to be typical rack-hanger packaging, complete with HyperX sticker for your case, if you are so inclined. :smile:

It is running quite well in P95. I have done some Memtest, mainly to see if it crapped out instantly or at least early. ATM, it is running Blend torture test with 15.2 GiB in use overall out of 16 GiB.

I ran it briefly in P95 DC, and it seems to live up to my hopes.

First, here results I can tie to the first pair of 2133, XMP 2400 (dual rank) DIMMS, running DC at 2400 MHz.

Then, there are the best results I could squeeze out of the single rank Corsair at XMP-3000, with the CPU at 4200 MHz instead of 4000. This was pretty much in Clean Boot mode to limit competition.

Finally, there is a quick grab of the current RAM at 2666 MHz with the CPU at 4000.:tu: (...on a larger exponent! :smile:) [with Windows in its typical loaded up state.]

Kingston 2666M dual rank +GTX1060 results
 

attachments:
[I][U]Kingston 2667M.JPG[/U][/I] shows the RAM at its sole published speed.

[I][U]start P95.JPG[/U][/I] shows the change in the change in performance of the 1060 in mfaktc when Prime95 is started, as indicated by the red mark. With previous GPUs (580), and CPUs (AMD FX) starting P95 caused a noticeable [U]decrease[/U] in G-D/D. Here, mfaktc speeds up.
[I][U]
hyperX-2666 at 3000.JPG [/U][/I]shows a snap of P95, with the RAM running at 3000M, at one-clock relaxation of basic timings, and V[SUB]mem[/SUB] bumped from 1.20 to 1.22 V.
This setup ran Blend torture test for ten hours. The RAM amount was set to 13 G (out of 16). FFTs were set at 2048 to 4096.

The final ms/it numbers are great!

Here's the latest tweak. Now the CPU is running at 4200 Mhz, up from 4000. The RAM is at 3000.