larrymoencurly's Discussion of Greater than DDR3-1600 RAM Speeds

[bmwbaxter]

Water? Ram? what? xD
I want submergible ram, with cl5 capable ic's and rated for 1600mhz by JEDEC, cause that's the only standard I trust and not those greedy guys at Crucial, which actually produce the less overclockable ic's these days lol

coat some hypers in vaseline and run them subzero. that should do the trick nicely

 

[ivanlabrie]

Or BBSE...vaseline is not exactly water, but condensation or frost are lol

 

[Bobnova]

How bout ~2646-5-9-7? http://img.hwbot.org/u28581/image_id_797005.jpeg
You can bet that neither the chips nor the sticks are rated for that

G.Skill 1600MHz sticks.

 

[ivanlabrie]

Speaking of which, you got pm.

And, this thread has gone bananas...

 

[larrymoencurly]

Those > $1M testing machines come in both the wafer-probing versions and versions that test whole packaged chips or even finished memory modules.

Nope, price doesn't matter in the slightest. Only accuracy matters. Why would price matter? We only care about accuracy. Guess what? That's all they care about too.

If the $2,000 testers are just as accurate, then the module companies that buy the >$1M machines are foolish spendthrifts, right? Some companies sell both modules tested by the expensive machines and by just PCs, and they charge extra for the former. How could they do that if it didn't matter, especially OEM customers? Neither question has been answered here.

Actually they do test for max clocks and tightest timings because the only way to bin the chips is by is by running them at the fastest specs and seeing which fail, then take the failures and test them at slower speeds, etc. I suspect module makers do the same, especialy because they use so many chips that are rejects or overclocked and don't know how well those chips can do. It's probably only the final testing of the labelled modules that's done at fixed, preset timings.

Do they? Have you ever personally binned a stack of RAM? Do you actually know how the process works? Have you binned anything?

Nope. I'm just going by what was told to me by an ex Motorola engineer where I work and, much more recently, in forums by other engineers who've been a lot more critical of overclocking than I am.

What fastest specs? I smell a lack of real knowledge here. Got any credentials, at all? Guess what: The people you're arguing with have done rather of a lot of binning and speed testing, of a lot of things.

Are you saying companies first test the whole batch at the slowest ratings and then ramp up to weed out the slower chips until the fastest speed rating is reached, rather than start at the fastest and going slower? If so, that makes sense, but otherwise could you explain?

Now let's move on the the stick itself.
Testing a RAM stick to the same level of known-stability consumes an 8th to a 16th of the time it takes to test the chips individually, they're testing 8 or 16 chips at once after all. Also there are far fewer sticks being tested, no memory stick manufacturer makes nearly as many sticks as the big fabs make chips!

For the opposite to be true, wouldn't each module have to average fewer than 1 RAM chip?

You make no sense, at all. Binning a stick with 16chips on it for 1h takes a 16th the time binning each chip individually for 1h takes. Basic math.

OK, I confused any single module maker with all module makers together.

A given design IS tested for max clocks at this point, probably a few sticks from that design. If the company is looking to make a fast stick of RAM they use the highest rating that most or all of the test sticks can run at with good stability.

Odds are decent that the RAM chips spend more time being tested at this point than they did in the factory.

How much of that is because they test more thoroughly than the chip makers do, and how much is due to them using slower testing machines?

Who said anything about machine speed? Did I? (cliffnote: Nope)

I did. Why are you asking? Longer test time doesn't necessarily mean the testing is more thorough, and that's why I wanted to know the speed of each tester.

I'm guessing more testing time at this stage because it is R&D, R&D takes forever on almost everything. Ever done any R&D? I'm guessing no.
Further, there is a bus limitation when you're testing a module, oh and you have to test the actual module too of course.

Doesn't it take more R&D to produce individual RAM chips than whole modules, especially when so much of the module industry shares board designs (IEEE or BrainPower)?

Once that's done, the production starts and each stick is tested for THAT set of speeds/timings, something that the chips themselves were never tested for. Had they tested the chips at higher speeds they could have rated them there, but they didn't.

Is that better than using chips guaranteed to meet those timings?

I'm Nope. Not worse, either. The end result is identical. I cite you for attempting to dodge the question.

False citation.

I could believe it if such binning actually was just as good, but it's not even close, according to the failures I've seen, and you've provided no evidence otherwise, especially with your not particularly credible claim that testing done by cheap machines is just as good as testing done by far more expensive machines, something even one maker of cheap testers said isn't true.

The bottom line is, if a stick is rated for 2133-7-10-7-27-1t at 1.65v, it will run that speed error free. If it doesn't, it's dead.

What about modules that consistently test fine during the first 15 minutes of computer operation but fail later? Some people will replace them, others will just raise the voltage or slow the timings and continue to use that "dead" memory.

What about those? They happen with everything. I've had genuine straight from intel CPUs die that way (actually only made it 5m at 100% stock, wasn't even out of the BIOS).

You said a module that didn't operate error-free at its ratings was dead. Anything that works even partially isn't dead.

BTW do also you think "decimate" means "completely destroy"?

Is that for 680K and 655K chips from the same production run? Batches vary in quality, even when clean room temperature is regulated to within 0.5 Celcius. It's illogical to compare speed bins from different batches, unless it's known that the batches are consistent enough with each other.

That's across every batch ever produced. If you're following that logic, it's illogical to believe that every RAM chip off a production line geared towards "1600MHz" chips will be correct.
You're arguing against yourself here, might want to rethink that one.
The variance of batches is decreasing as well, in the Intel CPU world batches have meant absolutely nothing for the last two generations.
Do you know what a batch is? It's more than one wafer, plenty of time for the temperature to change. Chips vary even on a single wafer, let alone per batch.

You can do testing that makes it highly, highly unlikely that any chips fall short of any parameter.

A batch is production run of wafers sliced from the same ingot, and I think yields over 90% are expected now, compared to as little as 15% in the 1970s for 16Kb or 64Kb Japanese DRAM.

If the chips are from different batches, why wouldn't the ones graded fastest be more likely to overclock better than the slower grades? The usual quality distribution curve for any one grade of chips should make the odds better for the fastest grade to overclock best and not favor the slower ones.

I ask again, where is your proof? Where is even a hint of proof? So far I've seen none, at all.

I haven't seen any from you, either, just a couple of isolated anecdotes. It's one thing to just mention such stories but another to generalize too much from them.

How can a brand's quality be good if its manufacturer's standards have allowed up to 2 faulty bits per module?

Every memory manufacturer I've dealt with was happy to RMA modules with even a single error in memtest86+.
This is a wild claim that demands some form of backup beyond "It's true cause I've said it a few times".

I don't know if I PMed the company's e-mail to you, but some other people here have seen it.

I've never had problems getting dud memory replaced, except in the case of a CompUSA PC133 module some Bible guy tested by walking behind a wall then returning in 15 seconds and pronouncing the module to test good.

Then roughly 10% of the sticks are bad, and almost always those sticks made from no-name or overclocked chips. Probably a lot more than 0.1% of the buyers have to spend extra time or money getting them replaced.

Any backup for those numbers that are spit out as factual?

I admit I don't have a formal database of the statistics and an analysis, but that's been about the dud rate for about the past 10-15 years with modules tested by MemTest86 and usually also Gold Memory, and except in one case, all of the failures involved no-name or overclocked chips. The one exception was that CompUSA module -- Infineon chips but rejected for solder on the contacts.

How many modules from how many kits have you personally tested? How many high speed modules have you personally tested? Do you even have a system capable of testing high speed RAM?

l think around 1,000 by now, but probably none of that's been high speed. All I use are PCs, but after the first overnight I stuff anti-static bubble wrap around the modules to make them hot. That's why I'm surprised at the high failure rate -- I'm not using any special tester or overclocking or undervolting.

Nobody who has actually tested high speed RAM is agreeing with you, have you noticed that?
Nor is anybody who has reviewed RAM agreeing with you.
Nor is anybody who knows anything about RAM agreeing with you.
Of course, all of the above could be because nobody is agreeing with you. Quite probably because you're wrong.

Where can I find reviews of RAM that aren't based on hand-picked samples sent out for free by the manufacturers? If they're anything like Kingston, they have modules that have been specially screened for defects, like the ones they RMA out to fussy customers (I've never gotten a dud RMA from them, even for the DDR line where 60% of the modules had errors). I have seen a compative overclocking review where modules made by a real chip company overclocked quite a bit better than others, but that's just one old test (DDR2), and I don't want to absolutely conclude anything from it.

In conclusion: Try backing up your actual statements rather than going off on wild attacks on the questions of your statements.

Show me one of my wild attacks, here or elsewhere.

 

[moocow]

l think around 1,000 by now, but probably none of that's been high speed. All I use are PCs, but after the first overnight I stuff anti-static bubble wrap around the modules to make them hot. That's why I'm surprised at the high failure rate -- I'm not using any special tester or overclocking or undervolting.

Best to just let this die guys instead of arguing when there is no point.

 

[bmwbaxter]

Best to just let this die guys instead of arguing when there is no point.

well played

 

[Jagged_Steel]

What you quoted wingman is downright hilarious to me. He acts as if Sammy return rates are exponentially better than other vendors. Truth is nobody knows. I GUESS however that they are all within a couple % of each other and all well under 10%.

A little known fact is that 99.9999999317183% of statistics quoted in internet forums are made up.

* I arrived at this absolutely provable number by averaging the number of statistics quoted in posts I have read divided by my interest level in them and then multiplying that by the number of flamewars that then ensued....

 

[Mr.Guvernment]

http://www.anandtech.com/show/4503/sandy-bridge-memory-scaling-choosing-the-best-ddr3

I think we confirmed what we pretty much knew all along: Sandy Bridge's improved memory controller has all but eliminated the need for extreme memory bandwidth, at least for this architecture. It's only when you get down to DDR3-1333 that you see a minor performance penalty. The sweet spot appears to be at DDR3-1600, where you will see a minor performance increase over DDR3-1333 with only a slight increase in cost. The performance increase gained by going up to DDR3-1866 or DDR3-2133 isn't nearly as pronounced.

 

[wingman99]

l think around 1,000 by now, but probably none of that's been high speed. All I use are PCs, but after the first overnight I stuff anti-static bubble wrap around the modules to make them hot. That's why I'm surprised at the high failure rate -- I'm not using any special tester or overclocking or undervolting.

Seriously What the heck are you talking about.

 

[Bobnova]

No clue really, no clue.
A lot of flag waving mostly, with some nonsense. The PM is hilarious by the way, it includes the date that G.Skill replied saying that they allow zero errors.

To the OP: You are the one trying to prove something here, the rest of us (all of us...) are saying there is NO reason to care what the ICs are rated at. You cannot prove a negative, basic logic.
You're saying the IS a reason, that's provable. Have at it, but that of course requires.... proof!

 

[Archer0915]

No clue really, no clue.
A lot of flag waving mostly, with some nonsense. The PM is hilarious by the way, it includes the date that G.Skill replied saying that they allow zero errors.

To the OP: You are the one trying to prove something here, the rest of us (all of us...) are saying there is NO reason to care what the ICs are rated at. You cannot prove a negative, basic logic.
You're saying the IS a reason, that's provable. Have at it, but that of course requires.... proof!

Bob a K cpu is rated for stock speed. A BE CPU is rated for stock speed. My Grand Marquis is rated at the legal speed limit.

So I don't run my CPUs at stock and other than blowing them up they perform flawlessly! And the purring V8 sees much more than the speed limit.

Some people get in in their heads that they know more than people with hundreds of years of collective experience. They do not understand what compliance is (yeah the memory will meet a standard and get a stamp or JEDEC members can cause them issues.

 

[RGone]

I will not post any of what I started to post. It was funny to me when I was thinking of what to write. Upon reflection, I think I will pass. There is nothing funny about this thread, for certain, by page #4. Wasn't the original LMC a comedy team?

 

[larrymoencurly]

No clue really, no clue.
A lot of flag waving mostly, with some nonsense. The PM is hilarious by the way, it includes the date that G.Skill replied saying that they allow zero errors.

IOW nothing I said publicly was contradicted. You wanted proof, I gave it to you, but you rejected it.

Again, if memory made with unbranded or overclocked chips was just as good, it wouldn't have failed so much more than other memory.

 

[ivanlabrie]

And it just goes on...buy whichever ram you like, and more than one set in case it fails on ya and cause you to spend money shipping it for rma and be done with it.
I will continue using g.skill 'overclocked' ic's myself

 

[Archer0915]

IOW nothing I said publicly was contradicted. You wanted proof, I gave it to you, but you rejected it.

Again, if memory made with unbranded or overclocked chips was just as good, it wouldn't have failed so much more than other memory.

So can you tell me the JEDEC specs for DDR3? What is the JEDEC spec for for DDR3 1866?

Did you know that 1333 and 1600 and all other DDR3 modules have the exact same Absolute clock HIGH pulse width and Absolute clock LOW pulse width?

Do you know what that is? What about jitter? Did you know the faster the memory the less jitter allowance?

Here are some specs:
Cumulative error across 2 cycles tERR(2per) - 147 147 - 132 132 - 118 118 - 103 103 ps
Cumulative error across 3 cycles tERR(3per) - 175 175 - 157 157 - 140 140 - 122 122 ps
Cumulative error across 4 cycles tERR(4per) - 194 194 - 175 175 - 155 155 - 136 136 ps
Cumulative error across 5 cycles tERR(5per) - 209 209 - 188 188 - 168 168 - 147 147 ps
Cumulative error across 6 cycles tERR(6per) - 222 222 - 200 200 - 177 177 - 155 155 ps
Cumulative error across 7 cycles tERR(7per) - 232 232 - 209 209 - 186 186 - 163 163 ps

 

[wagex]

did you know that if you put some memory in your computer it does stuff.

 

[EarthDog]

Ok, so I just recalled that someone, which happened to be me, formally posted actual return rates for products from some PC company overseas. Clearly these numbers are not as accurate as the actual company's doa/return rates, however its likely much larger of a data set than LMCurly's personal experience...(not sure I buy that you have tested 1000 kits, personally).

I recall saying above:

He acts as if Sammy return rates are exponentially better than other vendors. Truth is nobody knows. I GUESS however that they are all within a couple % of each other and all well under 10%.

Just so we are clear, the underlying meaning there is that is it worth it for you, LMC to insist to members, new members that are likely green behind the ears, to warrant buying a different brand/by IC's because:

I don't want to pay to ship RMAs, especially multiple RMAs, and it's easy to avoid defects in the first place..

You assertions and general tone, LMC, lead me, new people you 'helped' and the rest of the participants in this thread, that there is a significant enough difference between brands that you should buy by IC to prevent RMAs. So without further musings by me.. I present the best set of facts we can get our hands on.

Source thread: http://www.overclockers.com/forums/showthread.php?t=717026&highlight=Failure

- Crucial 0.23% (as against 0.4%)
- Kingston 0.40% (as against 0.5%)
- G.Skill 1.10% (as against 1.4%)
- Corsair 1.44% (as against 1.6%)

Crucial keeps top spot and, like the other manufacturers, improves its rate in a welcome overall downwards trend. Just two models have rates of over 5%, compared to 8 last year:

- 12.87% Corsair Dominator 4 GB (2x 2 GB) DDR3 1333 MHz CL9 - TW3X4G1333C9D
- 5.53% Corsair XMS2 2 GB DDR2 800 MHz CL5 – CM2X2048-6400C5

The worst Crucial model has a rate of 1.25%, the worst Kingston 2.1% and the worst G.Skill 3.14%.

What I take away from this is exactly what I said. There is ~1.2% difference between the best and worst vendors here with the worst vendor sporting a 98%+ success rate (not DOA or returned).

Another piece of information I thought you may find interesting LMC is all models combined and their return rates:

- RAM: 0.88% (against 1.20%)

All ram combined that was sold by this company were not returned OVER 99% if the time WITHOUT KNOWING THE IC.

One other thing I would like to point out is Corsairs stunning 12.8% return rate on the dominator 4... LMC, did you notice those are 1333MHz sticks? I bet those IC's are actually rated for 1333MHz and are not 'overclocked' (though that is a guess). Wouldnt that right there bunk nearly all of which you are saying and at the same time support other's assertions that its simply a difference of failure rates regardless of IC? I think it does.

So I ask again, do you think its worth all the trouble/discussion over a MAXIMUM of 1.2% difference? Nobody in their right mind would think that is a significant value...significant enough to push a 'choose by IC or face the much higher potential of RMAing' agenda across multiple threads which only served to confuse the OP's. Again, you may have saved people from a MAXIMUM RMA possibility of 1.2% (best vs worse)... Congratulations.

 

[azanimefan]

just to keep facts in this discussion.

Corsair Dominator TW3X4G1333C9 -
9-9-9-24 1.6V
DDR3 1333

Looks, oddly enough like it's overclocked, as the 1.6V would seem to indicate. Not 100% sure on that score, but from what i've read about it this is 1066mhz ram at 1.5v overclocked to 1333mhz and incredibly unstable at that speed... (apparently it even indicates these numbers on the chip itself) so they really did simply overclock some slow ram, put a heatsink on it, and called it "fast gaming ram"... cheap ram is cheap.

 

[EarthDog]

So be it. The overall point still remains...if its bad return it. Nobody needs to know the ic, period.