A Strategy of Options for Testing RAM?

[BonzaiDuck]

Nowadays, I always think it's a better avenue for progress to simply buy the RAM spec'd for the speed you want.

Testing RAM takes time, unless you run multiple instances of HCI Memtest-64 in windows. I'm more inclined to use the bootable-CD version without windows.

Here are some thoughts.

The RAM-makers spec their kits; some supposedly test them before they go into the packages for shipping out the door.

On the customer end, some user-reviews complain here and there of "DOA" or "failed after so many weeks/months."

The RAM-makers may exhibit great performance in RMA service; they may not. G.SKILL, for instance, is "great."

Suppose you want to test a new 2-module kit? The bigger the RAM capacity, the longer to test.

HCI Memtest suggests that a single-pass "100%" coverage is a "basic test." For a "thorough test," they recommend 10 passes or 1000%.

1000% for a 4x4 or 2x8 16GB kit on my SB systems can take four or five days!

But you bought these modules -- whatever brand. They're brand-new. You want to get past the possible defect, which -- if discovered -- means return for refund, replacement, or RMA replacement if you skipped past the reseller's window.

I'd say -- if the system boots to Windows at spec settings on a system proven stable otherwise -- 200% coverage should be adequate to satisfy yourself. But I may run 300%.

Now, suppose you decide to overclock these modules? There are different levels of "overclocking." You can change the timings; you can change the speed; you can change the voltage; you can change the command rate to 1.

So if you "found" what seems to be a functioning overclock, the ball is now in your court. You have strayed from the factory specs, unless you are underclocking -- an equally viable option in some cases. But in this situation, you definitely want a more thorough test. Is 500% adequate? Or does one need to go the limit and run 10 iterations?

And I'd think, for instance, if you simply want to change the command-rate, run at spec voltage, timings and speed, that's going to be a minor change. If it falls short, you can adjust the voltages (perhaps VCCIO as well as VDIMM) a tad, and it should probably clear things up. And I'd think that 400% or 500% would be adequate in that situation.

On the other hand, what would you do for bumping DDR3-1600 to DDR3-1866 with looser timings and the same voltage? 500%? 1000%?

 

[Nec_V20]

Nowadays, I always think it's a better avenue for progress to simply buy the RAM spec'd for the speed you want.

Testing RAM takes time, unless you run multiple instances of HCI Memtest-64 in windows. I'm more inclined to use the bootable-CD version without windows.

Here are some thoughts.

The RAM-makers spec their kits; some supposedly test them before they go into the packages for shipping out the door.

On the customer end, some user-reviews complain here and there of "DOA" or "failed after so many weeks/months."

The RAM-makers may exhibit great performance in RMA service; they may not. G.SKILL, for instance, is "great."

Suppose you want to test a new 2-module kit? The bigger the RAM capacity, the longer to test.

HCI Memtest suggests that a single-pass "100%" coverage is a "basic test." For a "thorough test," they recommend 10 passes or 1000%.

1000% for a 4x4 or 2x8 16GB kit on my SB systems can take four or five days!

But you bought these modules -- whatever brand. They're brand-new. You want to get past the possible defect, which -- if discovered -- means return for refund, replacement, or RMA replacement if you skipped past the reseller's window.

I'd say -- if the system boots to Windows at spec settings on a system proven stable otherwise -- 200% coverage should be adequate to satisfy yourself. But I may run 300%.

Now, suppose you decide to overclock these modules? There are different levels of "overclocking." You can change the timings; you can change the speed; you can change the voltage; you can change the command rate to 1.

So if you "found" what seems to be a functioning overclock, the ball is now in your court. You have strayed from the factory specs, unless you are underclocking -- an equally viable option in some cases. But in this situation, you definitely want a more thorough test. Is 500% adequate? Or does one need to go the limit and run 10 iterations?

And I'd think, for instance, if you simply want to change the command-rate, run at spec voltage, timings and speed, that's going to be a minor change. If it falls short, you can adjust the voltages (perhaps VCCIO as well as VDIMM) a tad, and it should probably clear things up. And I'd think that 400% or 500% would be adequate in that situation.

On the other hand, what would you do for bumping DDR3-1600 to DDR3-1866 with looser timings and the same voltage? 500%? 1000%?

I agree with your first sentence wholeheartedly.

Here's why. I had been running my new system on 16GB Patriot Memory Viper PC1600 RAM for the past year, I then decided to make a new "new" system.

The reason is that I saw two really sweet deals
1) Mobo: Gigabyte GA-Z97X-Gaming G1 WIFI-BK (normal cheap price £320 which I saw on offer for £191.75).
2) 32GB Corsair Vengeance Pro 2400 (4*8GB modules) (normal cheap price £262 I saw on offer for £160). BTW this deal is still available as of writing this available at DABS.

I don't know if you will agree with me or not, but in my experience increasing the performance of RAM gives you by far the least subjectively experienced increase in system performance. Sure you might be able to measure it on a specific benchmark, but from my own knowledge I know that I certainly don't feel it as much as say getting a faster GPU, CPU or replacing my HD as a boot disk for an SSD.

I have gone from PC1600 to PC2400 RAM and I cannot honestly say that I feel much if anything of a performance increase. The previous board I had in the system was also a good one (Gigabyte GA-Z87X -UD5H).

Bumping a PC1600 RAM to 1866 more often than not does not involve "looser timings" but loser timings. There isn't any point to bumping 1600 to 1866 when you have to cripple the timings to the extent that the effective performance will be less than if you had left it at 1600 in the first place.

With components failing I always go back to the PSU. If a person has bought a PSU as an afterthought because they have blown their budget on all the sexy bling and that cheap PSU has horrendous ripple (which they nearly always do), then don't expect components to last never mind overclock stably.

 

[BonzaiDuck]

We're pretty much on the same page.

Here, I'm examining the patience-trying troubles we have with stress-testing and proving stability. Like I said, you want the new RAM to meet its own specs just after you buy it. And also -- already said -- don't know about Haswell or '-E" -- but with the Sandy CPUs, it has taken me close to 5 days to run that "thorough" 1,000% test.

True, I only felt it necessary when I had a 4x4 kit of Ripjaws 1600 -- OC'd to 1866. While that was a "winner" for upping the CL from 9 to 10, setting the CR to 1 and leaving the voltage at 1.5, it was a lot of "down time."

With this trouble stressing over the AVX2 instruction set extension, the smaller die-size and higher temperatures, it now seems that OCCT:CPU, Intel XTU and AIDA-64 are said to be "validated" stress-test programs for stability, but they don't put the thermal burn on the chip like the traditional tests. Especially, OCCT:CPU boasts of finding errors in less than 4 hours if there are any to find, but folks still think it necessary to run Prime95 for 24 hours.

Throwing RAM into the testing regimen, one wants to know everything is tip-top, but without consuming a week's time. So -- yeah -- buy the RAM at the speed you want to run it, and if you're going to OC, it's going to take you . . . days. Unless, of course, insufficient time and effort is "OK." But it's not . . .

 

[Soulkeeper]

10 runs of memtest86 is redundant imo
If that's what you wanna do I won't fault you.

here's what I do:

1) 2-3 runs of memtest86 (include the hammer test!)
2) compile FF from source
3) play a video game for an hour
4) use the system daily

If it gets thru all that, then i never have an issue with it
compiling code has always been the best way to expose otherwise stable cpu/mem for me, i've relied on it for years.