Original Link: http://www.anandtech.com/show/2799
DDR3-2000+ Memory Kits - Fast but Flawedby Rajinder Gill on July 8, 2009 12:15 AM EST
- Posted in
We just looked at i7 memory scaling performance and now it is time to chill a few processors to see what those DDR3-2000+ kits are capable of for the serious overclocker. As it turns out that was the opening to our original article, which we planned to launch in conjunction with the DDR3 memory-scaling article.
However, the best plans of mice and men sometimes go awry. Unfortunately, we met delay after delay as every one of our Elpida “Hyper” based kits failed on us in some form or fashion over the past few weeks. At times, a single module would fail and eventually the whole kit in certain instances. Eventually our patience wore thin as even warranty replacements started failing and we knew this was not an isolated problem.
In fact, this problem has become widespread in the extreme overclocking community. Admittedly, widespread in this particular group means a few hundred users. Nevertheless, this audience purchases these extreme memory kits with prices tags reaching the $500+ level at times and expects like performance and quality in return. Certainly, the performance is there, quality we are not so sure about right now.
We could attribute the demise of our modules to the elevated voltage levels we have used for this article and normally we would go that direction and stop for the day. However, we had modules die on us using no more 1.50V VDimm and stock VTT settings in a variety of boards. We are not the only ones, as it seems a number of users have also been through the RMA process (a few more than once) regardless of voltage settings.
The “official” cause of death is unknown at present, while the usual suspects, such as manufacturing errors, motherboard voltage/ user over voltage issues and temperature related deterioration are the obvious perpetrators. The “unofficial” cause of death is simply a quality problem with the Elpida “Hyper” based ICs according to various sources we have spoken with the past couple of weeks. Granted, the other factors can and probably do account for a certain failure rate, but the randomness of our failures along with others, especially at first POST or during stock benchmarking lead us to believe that the quality of the IC is the primary factor at this point.
We have contacted Elpida about the problem but do not have an “official” response from their engineering group yet. However, the problem is serious enough that Corsair informed us earlier today that they asked their retailers to return any kits in the channel. They will not be selling kits based on the Elpida Hyper ICs until an enhanced manufacturing and testing process is in place to ensure the quality of this particular product before shipment. We applaud Corsair for being aggressive in regards to this problem and we expect/hope other suppliers to follow suit.
Most of the suppliers have reported that a relatively small percentage of kits appear to be affected. Just how small is unknown. Based on our own numbers and those of other users it appears to us it is significantly more serious than we were lead to believe a few weeks ago. We are now at the point of just saying that you are better off avoiding Elpida Hyper kits due to the ‘frequently random’ level of failures with modules. However, at least for now, all of the suppliers are fully backing their warranties. If you need the available benchmarking performance generated by these kits then it is worth the risk. Just make sure of the warranty terms before purchase, or simply put, buyer beware.
When we speak of failures, there are two types, a catastrophic failure where the module dies instantly and one of deterioration. One or more of the modules failing to map fully to the operating system usually marks the first sign of deterioration. Moving the modules around between the slots can work around some of this, although from our experience this is a primary sign that things are beginning to go downhill fast. This phenomenon is not to be confused with the i7 memory controller skipping to map a module because of insufficient voltages for the applied clocks.
The next step is when the module no longer clocks up at stock voltages or given voltage limits like 1.65V VDimm. We have witnessed modules not clocking above 1900MHz or so regardless of voltages and slowly dropping to 1200MHz before total failure.
With all that said, we decided to complete the article as there are users out there that have not experienced any of the issues at all or those still wishing to take a chance on these kits. Two of the kits we have been pushing for raw bandwidth over the past few weeks are Corsair’s Dominator GT 7-8-7-20 6GB kit and OCZ’s Blade 2133 8-9-8-24 6GB kit. We have tested other kits, but these are the last two standing in the labs although both of them are now showing signs of deterioration.
Corsair chose to stick with the tried and tested 2000MHz CAS 7-8-7-20 formula while OCZ gives up the tighter CAS rating in a bid to woo the market with a 2133MHz kit at 8-9-8-24 timings. On the face of it, we would say that both kits should be capable of similar results assuming that SPD and PCB differences between the two are not massive. OCZ and Corsair tell us their respective top end kits represent the top 1% of the Elpida Hyper yield.
We’re keeping it simple today, running a few of the preferred benchmarking programs in a bid to find maximum frequency limits for the modules along with a small comparison of scores at the same CPU frequency. We have already shown that these kits are not really needed by the general enthusiast and typically do not improve application performance significantly enough to warrant the increased cost. However, they do provide a certain degree of flexibility when overclocking and allow for very tight latencies at a variety of memory speeds.
Test Bed Setup
Windows XP is still the preferred operating system for 90% of the synthetic and 3D programs when benchmarking. We also utilized Vista 64 Ultimate to confirm memory stability in some of the 3D benchmarks with the entire memory footprint mapped and available. This allows us to find bandwidth limits at heavier system loads than those imposed by Super Pi under XP Pro.
We ran each benchmark five times with the high and low scores tossed out and the remaining scores averaged. We utilized fresh operating system installations for each memory kit. We reboot the system between each benchmark program change, clear the prefetch folder, and then defragment the drive between reboots to ensure the results are consistent.
We’ve got a choice of 3 sub-zero temperature friendly CPU’s in the labs (all D0 stepping). Surprisingly enough the 975 D0 taps out earliest, and just about crawls to 5GHz providing the QPI link frequency multiplier is left at the lowest possible setting. We had expected this processor to give us the requisite flexibility of pushing these kits to the pinnacle of their ability, but the i7 920 and W3540 Xeon proved to be the better choice for raw bandwidth on both sides of the bus.
We matched memory sub timings on both modules, right down to Round Trip latency parameters in a bid to ensure that maximum frequency scaling is not hindered by incorrectly sensed values. EVGA’s Classified E760 motherboard provides BIOS level overrides for almost all memory parameters as well as auto presets for all available sub-timings based on SPD information.
CAS Fashion Question: Is CAS 8 the new CAS 7?
We’ll start off with Super Pi to check frequency scaling. We set a memory voltage ceiling of 1.85V, which we think is more than most users will ever push into their Hyper based modules for benching. For 24/7 use, we do not suggest more than 1.72V based on conversations with the memory suppliers, actually less is more in this case.
The Corsair modules offer just slightly better voltage/frequency scaling in the Super Pi results. We attribute most of the differences here to the variances in the kits we received. The original Corsair kit actually generated slightly lower results when it came to maximum memory clocks. Our first Blade kit was marginally better when it came to voltages at like settings. The kits tie where the CPU is the limiting factor once we are over 2300MHz with CAS 8-9-8 timings.
Peak gains when using a higher memory bandwidth are certainly apparent between CAS 6 and CAS 7. With the Uncore multiplier at 16X, CAS 6 on the 2:8 memory ratio loses around 10 seconds to CAS 7 running at 2094 MHz with an Uncore multiplier of 20X. In case you’re wondering why there are no Corsair results at CAS 7-7-7 timings, it because out of the six modules we have in the labs, none will post consistently over 1950MHz with CAS 7-7-7 timings at this point. This is irrespective of VDimm settings or swapping the modules around in the slots as we notice early deterioration setting in at this point. In fact, a couple of the modules are already near a 1500MHz limit.
Oddly enough 7-8-7 or 8-9-8 timings still work fine on several of the Corsair modules at 2000, provided we put the module into the right slot to initialize the POST process. This simply comes back to the modules deteriorating through the course of our testing. Our OCZ Blade modules can be plugged in at these timings and still POST although we have started to encounter a few module swap requirements. We are not ruling the board or voltages completely out of the equation just yet, but based upon previous testing results we do not believe the situation is specific to either variance as one of our OCZ modules already failed at stock voltages.
Even a matched Uncore multiplier at CAS 6 is not sufficient to eclipse the CAS 7 2094 MHz 32m time. The Everest results for these 2 test points do confirm the performance advantage of having higher bandwidth and 2ns faster access latency.
Perhaps more interesting than the raw gain versus memory frequency is something we found out during our tests of the EVGA Classified a couple of months ago. The light capacitive load of the Hyper modules allows higher overall processor clocks than memory kits based on Micron or Samsung based parts, reliability issues withstanding of course.
CAS 7 is likely to be the preferred latency setting for most benchmarks, providing enough flexibility for scaling processor clocks right to the limits of the silicon, especially when using locked multiplier CPU’s. Running either set of modules at CAS 8 to around 2250MHz does allow the use of slightly more conservative memory voltage at the cost of the low acess latency of CAS 7-7 timings around 2100MHz. We added a couple of maximum frequency runs above just to show how it all adds up.
A few 3D results...
Let’s take a quick look as to how 3D benchmarks such as 3DMark Vantage, 3D Mark 06 and 3D Mark 05 are believed to be somewhat affected by memory bandwidth. Vista 64 allows us to see how the memory voltages we utilized for Super Pi apply to a 64 bit OS when all 6GB of memory is fully mapped.
No major gains here in 3DMark Vantage as we have a 100-point difference between CAS6 1836MHz and CAS8 2294MHz.. Voltage wise, we see the need for an elevated idle voltage in Vista due to increased voltage droop under load with 6GB fully mapped. The actual load Vdimm in these benchmarks is exactly the same as what we used for Super Pi 32m under the XP operating system.
3D Mark 06 scores more consistently run to run than Vantage and seems to prefer the tighter CAS latency of 6, giving a 400 point boost.
Switching over to XP for 3D Mark 05 we find that the GPU benches are not overly sensitive to CAS or memory bandwidth. The best overall scores come when using the 20X Uncore multiplier on our 920 CPU (one below the CPU core multiplier). Using higher Uncore multipliers than the CPU core multiplier actually resulted in lower average results for 3D Mark 05, due to the additional hold times asserted by the IMC to the clock crossing process. This performance offset can be seen in BIOS via the increased Round Trip Latency values. In essence you get faster L3 cache performance while the IMC slows down memory bus transfer times to levy the increased performance on the other side of the bus; giving up one for the other. Bottom line here is that higher Uncore frequencies don't really offer anything exciting in any benchmark that reads or writes more data than the CPU cache can hold. As an example running a 23X Uncore multiplier results in a 15ns shift for the worse in Round Trip Latency which cannot be reclaimed with the current range of settings availble withn the EVGA BIOS. Higher Uncore frequencies may appear attractive at first approximation, but they're not always best in every scenario.
Using CAS 6 also resulted in the GPU tests of 3D Mark 05 failing to pass at 229 BCLK regardless of voltages, while running 2294 MHZ CAS 8 with the same Uncore multiplier passed without a glitch.
We had certainly hoped to provide a review with all of the latest DDR3-2000 6GB kits, but it was just not meant to be considering the failures encountered during testing. Based upon the frequencies and timings of the Elpida Hyper based modules from OCZ and Corsair, these kits offer nearly the same performance with voltage requirements favoring Corsair although this can be attributed to the variances in the kits we received.
We would love to recommend either one to any user who is fanatical about benchmarking or just having the best product in their system. Unfortunately, it is not something we can do based upon the current reliability issues. We accept that elevated voltage levels above 1.65V are out of potential warranty levels, even though Elpida designed the IC to scale with voltages up to 1.85V. However, the deterioration/failure of this particular IC seems to be far in excess of anything we have had in the labs for quite a while.
Based on our own experience the failures are not vendor specific as modules from all the major suppliers have died on us. We lost modules using as little as 1.50 VDimm, on five different Intel boards, an AMD AM3 board, and at various temperatures on the modules and IOH. These failures are not solely confined to a particular slot location, parts combination, or BIOS design, leaving little doubt that there is a quality assurance problem at Elplida. Whether that quality problem is based on a lack of proper screening, process controls, engineering design, or a combination of all three is something we may never know.
What we do know is that Corsair, OCZ, and others are honoring their warranties with Corsair now reaching the point of pulling their kits from the retail channel until these problems with Elpida are addressed. Whether other suppliers follow suit or not is something we are trying to verify at this point, although initial conversations indicate this might happen.
If you are shooting for world records, then there really is no other choice at present, it’s pretty much Elpida Hyper or bust when you look at top scores on the i7 platform on the Orb or HWbot. Of course, that does not matter if the kits fail on you during a benchmark run or are unavailable. We hear rumors that Samsung might have something in the pipeline shortly that will at least mimic the high bandwidth of the Elpida Hyper at low voltages. We hope that performance will come without the attendant reliability problems we are suffering with currently.
There is a glimmer of hope as a select few users seem to have a bit of luck and have had no failures at all, even after months of usage. The question is at $500 a pop for these particular kits, do you feel lucky enough to take the risk?