Unscientific findings with Kingston DDR3-1600

[jgreco]

There have been lots of reports that the Kingston DDR3-1600 product (specifically KVR16E11K4/32) hasn't been working with Haswell systems like the SuperMicro X10 boards.

We haven't had reason to order an X10 based board, but as some of you know we got an ASRock Avoton board in a few weeks ago. This board IS on the Kingston web site under their compatibility tool. This failed rather spectacularly with the set of Kingston DDR3-1600 we had in inventory as spares.

I decided I didn't want to bring another 32GB of memory into stock just to test, so this morning I downed one of our X9SCL+'s and stuffed the DDR3-1600 into it along with a gen 1 E3-1230 (which would run it at 1333).

Failing on two boards it is listed as working with. Dunno!

 

[Yatti420]

Have you tried fail safe mode etc? Also try the 420 version.. Oh brb..

 

[Z300M]

There have been lots of reports that the Kingston DDR3-1600 product (specifically KVR16E11K4/32) hasn't been working with Haswell systems like the SuperMicro X10 boards.

We haven't had reason to order an X10 based board, but as some of you know we got an ASRock Avoton board in a few weeks ago. This board IS on the Kingston web site under their compatibility tool. This failed rather spectacularly with the set of Kingston DDR3-1600 we had in inventory as spares.

I decided I didn't want to bring another 32GB of memory into stock just to test, so this morning I downed one of our X9SCL+'s and stuffed the DDR3-1600 into it along with a gen 1 E3-1230 (which would run it at 1333).

View attachment 3727

Failing on two boards it is listed as working with. Dunno!

And searching at Newegg.com for the 8GB "Server Premier" modules (KVR16E11/8EF) that Kingston used to recommend for the X10-series boards now results in a "not available" message -- not just "out of stock" but "not available." (You have to search by Newegg item number -- N82E16820239649 -- as searching by Kingston's Part No. finds nothing at all.)

Update: Searching for that Part Number on the Kingston Web site now shows it as "discontinued" and replaced by the KVR16E11/8KF (K for Kingston rather than E for Elpida) -- but even that is not listed as compatible with the Supermicro X10 boards.

 

[joeschmuck]

I recall reading that one of the MemTest86 programs would generate false failures. I don't recall the details behind it but with any luck that is what you are running into. I need to look into it myself because I suddenly have a failure on a system that tested fine before but I think I used the version of MemTest86 on the Ultimate Boot CD for the earlier test so I may return to that to verify the failure.

 

[Z300M]

I recall reading that one of the MemTest86 programs would generate false failures. I don't recall the details behind it but with any luck that is what you are running into. I need to look into it myself because I suddenly have a failure on a system that tested fine before but I think I used the version of MemTest86 on the Ultimate Boot CD for the earlier test so I may return to that to verify the failure.

I currently have MemTest86+ ver. 5.01 running on my X10SL7-F with two of the formerly recommended but now-discontinued KVR16E11/8EF modules. Six passes so far without an error -- but, from what Kingston tech. support told me a few weeks back, the problems occurred only when using four modules anyway.

Edit: Now at eight error-free passes.

My main concern now is what will happen if I want to add two more modules later: will the current modules be compatible with whatever else I am able to buy when the time comes?

 

[joeschmuck]

The issue was with certain configurations and had the issue of periodically popping up a random failure. 8 passes is quite a lot of testing, are you going for a specific number of passes, I think 3 should be good.

 

[cyberjock]

You know, the 3 pass thing has always bugged me. I mean, there's so many tests run(300+) per pass that normally I'd think that 1 pass would be fine. I'm not saying that 3 passes aren't the recommended number of passes, I'm just questioning where the 3 passes "thumbrule" came from and whether there's actually any engineering characteristics involved with coming up with 3 passes.

Now for an anecdotal story...

I was building a server for a friend and of course we chose to do a RAM test. The hardware was being repurposed so wasn't brand new or anything. Anyway, the unit was built and in the rack. RAM test ran for 3 passes with no errors(woohoo). Well, I never stopped it so it ran over the weekend while I was out of town. When I came back there was something like 150 errors. The first error was on pass 15 or so, and it was currently on pass 28 or so. In a more confusing twist, the errors all occurred between 15 and 25 passes or so, so there had been 3 passes since the last error was detected.

 

[jgreco]

Honestly, the three pass thing bothers me too. I have no idea why anyone would think a mere three passes would be sufficient to catch all error modes. As the scale of semiconductors has shrunk, the sheer accuracy required in the processes is mind-boggling, stunning, and totally amazing. That we can manufacture a device containing billions of bits with perfect accuracy seems incredibly unlikely. One would think that there might be a little leakage here or there. And that's what the memory tests are there to identify. If you study the actual tests, many of them are looking for problems with adjacencies and all sorts of other obscure but known failure modes.

Part of the problem is that kids that have grown up in the digital mindset forget that digital is a figment of your imagination ... electronics are fundamentally analog. In the 5V TTL world, for example, a zero is represented as zero volts (ground) and a one is represented at +5V DC. But what is a circuit to do when a 4V signal comes along? Will it be seen as a zero? A one? What if it's right at the threshold point? (p.s. I don't need any EE's trying to tell me how much I'm oversimplifying; I *know* I am.)

The average system has immense amounts of memory attached to it in this modern era. An E3-1230 has a max memory bandwidth of 21GB/sec. That means that at peak speed, it can only access each memory location on a 32GB system about 56000 times in a day. That's not a lot. You need to try so many things to identify even consistently repeatable problems, much less the rarer obscure issues that only strike every 1000 or 100,000 memory accesses.

 

[cyberjock]

Part of the problem is that kids that have grown up in the digital mindset forget that digital is a figment of your imagination ... electronics are fundamentally analog. In the 5V TTL world, for example, a zero is represented as zero volts (ground) and a one is represented at +5V DC. But what is a circuit to do when a 4V signal comes along? Will it be seen as a zero? A one? What if it's right at the threshold point? (p.s. I don't need any EE's trying to tell me how much I'm oversimplifying; I *know* I am.)

Have you read about how DVI and HDMI signals are processed? If not, you might want to check that out. It deals with this digital->analog->digital issue. ;) It literally reconstructs the original digital data at the "destination" despite the fact that digital signals are made "more analog" than most computing technology can handle.

I won't explain it since you don't want any EEs telling you how much you are oversimplifying. ;)

 

[Z300M]

The issue was with certain configurations and had the issue of periodically popping up a random failure. 8 passes is quite a lot of testing, are you going for a specific number of passes, I think 3 should be good.

I ended up letting it complete 10 passes. No particular reason for that many, except that I was out most of the day and did not need the machine to be performing its regular FreeNAS duties, so I just left it running.

 

[joeschmuck]

I don't know where 3 passes came from either, just what I've heard and been using for several years. Maybe we should do it longer.

 

[D4nthr4x]

I'm running 2x 8gb kingston so-dimms http://www.newegg.com/Product/Product.aspx?Item=N82E16820239702 in my supermicro a1sri-2758f avaton board. I ran through a pass and didn't see any errors. Kingston still lists this ram as compatible with this board on their site though.

 

[DataMover]

I'm running 2x 8gb kingston so-dimms http://www.newegg.com/Product/Product.aspx?Item=N82E16820239702 in my supermicro a1sri-2758f avaton board. I ran through a pass and didn't see any errors. Kingston still lists this ram as compatible with this board on their site though.

That's true for me, too: same board, same memory (KVR16LSE11/8). Honestly, I do follow to some extent all the hardware concerns.....but there is no error-free thing in the end. And it's for home use, at least for me. I haven't done an intensive (over weeks) disk test either.

My build runs. The Kingston memory works. Who can really explain what happens during weird memory tests down to the physical layer, regarding a special module, it's special used materials, atomic defects, ....