You just bought the fastest (and most expensive) desktop platform on the planet. Which company's memory will you use to populate Intel's quad-channel controller? We tested four purportedly high-end kits in order to find out which set is the best.
Anyone willing to drop one thousand dollars on a CPU probably wants to match it up to the best possible memory kit, especially armed with the knowledge that Intel's new quad-channel memory controller is located within that CPU. How are we to know which kit is the best, though?
We left it up to manufacturers to decide which quad-channel kit they thought would be best, making it clear that our judgement would be balanced between overclocking, minimum latencies, and capacity.
The only premium memory kit in today's round-up containing four 2 GB modules, Corsair's 8 GB DDR3-2400 kit promises top overclocking potential. These two things are probably related, since memory controllers are often capable of being pushed a little harder when paired with lower-density RAM.
As with other Dominator GT memory, the CMGTX8 kit features removable fins and Corsair's DHX Pro connector on each module. The fins can be replaced by parts of alternative size and color or, if you can still find one, the firm's long-discontinued thermoelectric cooler. The DHX Pro connector is designed exclusively to support Corsair's AirFlow Pro temperature and activity display.
Corsair skipped the now-standard DDR3-1600 C9 defaults in what appears to be an effort to assure the ultimate compatibility, since the lower speed will almost assuredly boot on nearly any board. Data rates lower than DDR3-1066 are no longer needed, since Sandy Bridge-E supplants the only processor architecture left officially constrained to that speed, Gulftown. We're not sure which platforms will treat the 518 MHz value as DDR3-1066, though.
CPU-Z doesn't report the correct frequency for XMP-2400, but our motherboard read it without issue and set the appropriate timings automatically when switched to XMP Profile 1 in its UEFI.
Corsair DRAM carries a non-transferable limited lifetime warranty.
With a model number that's easy to decipher but difficult to remember, G.Skill's PC-19200 CL 9 quad-channel kit is the only 16 GB set in today's round-up to carry a DDR3-2400 rating.
G.Skill says that, apart from its XMP profile, this is the same hardware as found in its DDR3-2200 kit. That made it the perfect product for a surreptitious appearance in our recent X79 motherboard round-up, where it was used as the second set for eight-DIMM overclocking tests.
We were happy to see G.Skill's DDR3-2400 automatically configured at DDR3-1600, but a little disappointed that JEDEC's slowest CAS 11 timings were used. While it is possible that G.Skill was simply seeking the ultimate compatibility, it's been a while since we've seen a "performance" motherboard that didn't support DDR3-1600 CAS 9 at 1.50 V.
Boards that can't use DDR3-1600 by default will find lower SPD values, while those that support XMP will configure the correct DDR3-2400 timings using Profile 1 from UEFI. CPU-Z misreports that value as a data rate of 2286 MHz, but our motherboard had no problem reading it correctly.
G.Skill memory includes a lifetime warranty.
Geil shipped its latest 16 GB quad-channel kit to us before any of its U.S. vendors received shipments. The unfortunate result is that its resellers have different offerings available to our North American readers. Versions of this kit at DDR3-1866 C9 and DDR3-2400 C10 are available here for $150 and $325, but this DDR3-2133 SKU can only be found in Europe (for the equivalent of $260). Reader demand could bring it to our shores, but the existing availability issue means that it's out of the award race.
Our motherboard detects and configures Geil's DDR3-2133 C9 using its DDR3-1600 C9 SPD. This, of course, is bootable at a motherboard's default 1.50 V, and yet Geil is the only company in today's round-up with the guts to add this configuration value.
DDR3-2133 CAS 9-11-9-28 is selectable as Profile 1 from XMP-compatible motherboards.
The least-expensive of this comparison's premium RAM, Mushkin's Redline 993997 still packs the punch of a DDR3-2133 rating and a full 16 GB capacity.
We're not certain how Mushkin achieved this pricing feat while using more-expensive heat spreaders than its 16 GB competitors. But we're happy to see that this kit is still as easy to configure as those more-expensive parts. DDR3-2133 CAS 9-11-10-28 is available as XMP Profile 1.
SPD profiles, on the other hand, are far more conservative. So, the best our board could accomplish without enabling XMP was DDR3-1333 CAS 9. Manual configuration is course another (highly recommended) option in this case.
Mushkin provides a non-transferable limited lifetime warranty to the original purchaser of its DRAM products.
We wanted to see what effect various memory speeds might have on program performance, and games are one of the types of programs that occasionally show this difference. Nvidia's GeForce GTX 580 is fast enough to keep the pressure on our CPU and GPU.
Intel's Core i7-3960X was locked at 34x throughout testing to keep its clock frequency stable at non-reference base clocks.
The lowest-possible game settings would show the biggest impact of memory performance on frames-per-second, but nobody actually games at those settings. Instead, we selected the lowest settings that high-end buyers would likely use (if forced to do so), along with a couple other applications that have been influenced by memory performance in the past.
Only two of the memory kits in today's round-up are rated at DDR3-2400, while others list DDR3-2133 as their top validated data rate. So how many kits are actually capable of pushing the high mark on our motherboard?
Corsairs 8 GB kit barely exceeded expectations, while RipjawsZ fell slightly below them. In spite of that, G.Skill's kit is still the fastest among 16 GB competitors.
Ergo our caveats, the first of which is that many Sandy Bridge-E-based processors can't reliably run at DDR3-2400. Our first Core i7-3960X wouldn't reach this frequency at any setting, and its replacement is barely any better in this regard (though its higher-attainable core speed does make motherboard testing easier).
The second caveat comes from Asus, which informed us that CPU VCCA (formerly referred to as uncore) voltage probably shouldn't exceed 1.20 V. According to them, some processors lose stability and may even be damaged at higher settings, while others do not or will not. Ours reached the highest memory settings using 1.20 to 1.25 V.
Our third caveat is that different memory firms use different CPU samples to determine the best settings for their own memory, and those settings might not be right for your CPU. Corsair, for example, included VCCA of 1.40 V in its XMP profile, but wouldn't overclock well until we dropped it to 1.25 V. Likewise, G.Skill's 1.20 V XMP VCCA worked perfectly with our CPU, but probably wouldn't work well with whatever CPU the folks at Corsair used to dial in their profile.
None of the DDR3-2133 samples include VCCA in their XMP profiles, likely because the CPU's memory controller only needs this adjustment at extremely high data rates.
Corsair has the best DDR3-2400 timings, though this is likely a result of its lower-density 8 GB kit. Our CPU sample, hand-picked for its overclocking headroom, couldn't handle the larger, higher-density 16 GB kits at this frequency. That also eliminates any DDR3-2400 performance testing, since we'd need at least two devices to compare.
G.Skill takes over from DDR3-2133 downward, its 16 GB kit pushing 1-cycle lower CAS at every tested frequency, compared to competitors.
One way to boost the overclocking capability of memory is to give it slower secondary and, if possible, tertiary timings. We'd hope that wouldn't hurt performance at lower data rates but, well…
Not interested in causing any panic, we went on to test the gaming performance of each module set. The slowest set in Sandra takes second-place here. Clearly, this platform isn't starved for memory bandwidth, else we'd see a closer correlation between the two sets of tests.
Similarly, Corsair ends up in a five-way tie that includes the reference DDR3-1600 CAS 9, with the lowest-latency G.Skill RAM squeaking ahead by a few milliseconds in 3ds Max.
We've seen that Sandra Memory Bandwidth doesn't always reflect the performance differences found in real-world applications, but we're still a little unsettled to find the 8 GB kit so far behind its 16 GB rivals at DDR3-1866. We're left to wonder how much module organization counts here.
Fortunately for Corsair, we again see that Sandra's numbers aren't reflected in real-world benchmarks. The other side of that coin is a realization that faster memory data rates simply do not translate to better performance. Desktop-class applications just don't tax a throughput-heavy quad-channel memory controller.
By now, most readers have noticed that the performance of DDR3-1600 reference memory doesn't change. That's because it's always at DDR3-1600 CAS 9, regardless of the data rate its on-paper competition is running. In this case, everything but the reference memory is operating at DDR3-2133, which should kick performance up a couple notches.
The slowest memory in Sandra actually leads our gaming performance charts, while DDR3-1600 falls noticeably behind only in WinRAR.
Click here to view gaming performance results.
Click here to view application performance results.
Corsair wins our overclocking competition, and that should be enough to convince competitive overclockers to take the plunge. Of course, in the real-world, you have to remember that it takes stepping down to an 8 GB kit to achieve those marginally-higher data rates. The rest of the field sports 16 GB, and we have to imagine that most buyers will be inclined to buy larger kits for expensive Sandy Bridge-E-based platforms.
Going the 8 GB route is an option, of course, if you're building on a budget. If that's the case, though, Corsair's premium is going to put it out of reach. Again, this one's probably best suited to the competitive overclocking circuit.
So, how do we define superiority in a market were people are willing to pay 100% more to get 10% better performance? Certainly we can't make a decision based on price alone.
After all, the second-fastest G.Skill kit has twice the capacity, and anyone who wants to argue price can point out that the 100% increase in capacity comes at a mere 26% increase in cost.
But consider also that Mushkin came up only 9 MHz-not nine percent-lower in maximum data rate. In other words, we got 99.6% of the most-expensive kits clock speed for only 33% the price. And for those who prefer not to split hairs, 99.6% rounds up.
Today's competition was presented as a complement to our recent high-end X79 Express motherboard round-up for the folks willing to spend top-dollar on Intel's Core i7-3000-series processors. The goal was to root out the best memory kit to go with Sandy Bridge-E's quad-channel memory controller, and it turns out that we have two bests. Corsair's Dominator GT CMGTX8 8 GB quad-channel kit reached the highest frequency, while G.Skill's RipjawsZ was the best overclocker among 16 GB kits. But mimicking (if not duplicating) the top 16 GB kits achievement at one-third of its price makes Mushkin's Redline 993997 the kit we'd recommend to our friends.