G.Skill is no stranger to pushing the limits of what RAM is capable of. Recently we showed you a DDR3-2666 kit they displayed at CeBit. Today we look at some of the fastest 8 GB DIMMs on the market – a 32 GB kit (4 x 8 GB) rated at DDR3-2133 that even comes with decent timings of 9-11-11-31-2T.
Specifications & Product Tour
The specifications are impressive, not just for the speed but also for the voltage. These high density sticks are rated at this speed while sipping only 1.6 V. They are advertised for use with the P67 / Z68 chipsets, but we will be examining them (with G.Skill’s blessing) on an X79 system in quad channel.
Model Number F3-2133C9Q-32GXH Series RipjawsX System DESKTOP System Type DDR3 Main Board INTEL M/B Chipset P67 / Z68 Capacity 32GB (8GBx4) Multi-Channel Kit Dual Channel kit Tested Speed DDR3-2133 MHz (PC3-17000) Tested Latency 9-11-11 2N Tested Voltage 1.6V Registered/Unbuffered Unbuffered Error Checking Non-ECC Type 240-pin DIMM SPD Speed 1333 MHz SPD Voltage 1.5V Warranty Lifetime Fan lncluded No Features Intel XMP (Extreme Momery Profile)Ready
Thankfully G.Skill has chosen to forego the fans this time around. They just plain aren’t needed. Unless you have an improbable (and unwise) fanless case with zero airflow, directly cooling RAM with extra fans is only for looks. If you do have a fanless case, you have much bigger problems than warm RAM. The voltages current-gen Intel CPUs can handle through their memory controller aren’t high enough to worry about fans for normal operation. If you need extra cooling, it will be in the form of liquid nitrogen to really pound on some sticks hunting down max frequency.
Thus, this kit comes in a clam-shell plastic, space-saving package. The sticks themselves are shrouded with G.Skill’s black RipjawsX heatsinks with blue accents, which makes for a good looking set of RAM.
Of course, I’m never content to just take a couple photos, so enjoy a plethora of fun angles!
Last but not least, a photo inspired by MattNo5ss’ review. Best RAM photo positioning ever.
Solid specs and a good looking heatsink to go with them, what’s not to like?
Our test system today is the same as that from our RipjawsZ 16 GB DDR3-2133 review.
|Motherboard||ASUS Rampage IV Extreme|
|RAM||G.Skill RipjawsX 32GB DDR3-2133/9-11-11-31
G.Skill RipjawsZ 16GB DDR3-2133 / 9-11-9-28
|OS||Windows 7 Professional x64|
It is housed in the massive, reviewer-friendly, easy product switching Lian Li PC-Z70.
For comparison purposes, we’ll be looking at the same comparisons from before, the RipjawsZ kit in dual-, triple- and quad-channel as well as that same kit running quad channel DDR3-1600, 9-9-9-24.
Stability at Rated Speed
Before performance testing, it’s always good to look at stability. Ultra-X has generously supplied their RAM stress tester, the R.S.T. Pro 3 PCI Express. There are only a couple of review sites that have one of these and we are proud to be one of them. It’s not your standard Memtest++ bootable CD. This is a standalone, bootable piece of hardware that plugs into a PCIe x1 slot and, man, does it ever beat on some RAM.
Several manufacturers use this to test their memory before it goes out the door to reduce RMAs. If it passes this test, it is stable and then some. Manufacturers generally run the full memory test for three passes. We’re overclockers though and tend to abuse sticks a little more than other people, so for our purposes, five passes sounds good. Running this test on 32 GB of RAM ended up absolutely hammering the RAM overnight and then some with a load more that it will ever experience in every day use, so that should be sufficient to say whether or not they’re stable.
One very important note is that these were set manually to their XMP profile timings. I prefer manual settings, but you could use the XMP profile if you wished. If you do not do set these timings, “Auto” will very likely be too tight on these sticks. In order to get decent primary timings, sacrifices had to be made down the line with sub-timings. We’ll see how that affects things when we bench them below, but please keep this in mind when setting these sticks up – you must set proper timings for stable operation!
Once the timings are set per the XMP profile though, as you can see everything is happy and stable through the R.S.T. Pro torture test. In case you want real-world stability testing too, I ran HyperPi with 12 threads and it, too, passed with flying colors.
We can unequivocally say this kit is stable at its rated speed.
Within reasonable constraints (i.e. if timings are relative to speed, getting tighter at slower speeds and vice versa), typical RAM kit performance variation is not necessarily going to be anything to call home about. It will show in synthetic tests and RAM-specific benches (SuperPi 32M), but less so with actual use in most cases.
The reason for the paragraph above is specifically the SiSoft Sandra testing. Those secondary timings definitely have an impact with synthetic benchmarking. You’ve got to make some sacrifices when you want to run such high density modules at this kind of speed and the sacrifices are evident here.
Memory bandwidth takes a pretty significant hit here. The quad-channel boost you get with Sandy Bridge-E is pretty much offset by the timing-for-density compromises, making this kit at quad channel similar to a dual-channel DDR3-2133 kit with 9-11-9-28 timings and tighter sub-timings.
Cache bandwidth takes a pretty significant hit, as does latency. Again, these tests are tailor made to show significant differences. You won’t see this type of variation in any real-world use.
Rendering, Compression & Encoding
Speaking of real world, when we do some render benchmarking with Cinebench, we see that variation is less than 2% across all tests.
7zip is a slightly different story. Compression seems to depend highly on memory bandwidth and those sub-timings did affect results, with the more losely-rated kit coming in about 7-8% slower than the 16 GB kit.
Now we’ll check out a little more rendering and video encoding performance. The former, via PoV Ray, shows it doesn’t give much of a rip about things, with top-to-bottom variation of 0.8%. x264 benchmark shows some difference in the reading pass (Pass 1), but when actually encoding video (Pass 2 – and the more important of the two), the difference drops back down within 1% top to bottom.
Save compression, whose variation definitely showed a difference with looser timings, real-world applications don’t seem to much care about the compromise to get this kit running as fast as it can.
WPrime historically doesn’t care one lick about RAM. There are members on our benchmarking team that run it with one stick in single channel and swear they get the same performance as dual channel. Indeed, benchmarking this kit with it bears this out with a maximum variation (and only in one instance) of 2%. Other than that one WPrime 32M result, variation is within 0.5%.
SuperPi is a different animal and it shows that benchmarkers won’t want to grab a 32 GB kit of RAM for pushing their CPU to its absolute limit for HWBot benchmarking. But you knew that already! Even so, anything that uses single-threaded floating-point calculations will show up to just under 7% performance loss due to the loosened timings.
Even with that difference though, single-threaded floating-point calculations aren’t going to be what people want to get this much memory for. People interested in 32 GB of fast memory do a lot of audio encoding, a lot of video encoding/transcoding/rendering, a lot of CAD, a lot of image rendering/manipulation, etc. etc. Basically, they’re going to do things that benefit more from the extra 16 GB of memory than they do the tighter timings on the smaller kit.
The first thing we’ll look at in our overclocking adventure is timings. One thing I enjoy doing, especially on kits with looser timings, is dropping the DRAM divider by one and seeing how tight the RAM can go. In this case, the next lower divider ran the RAM at DDR3-1866. Then all the primary timings were dropped from 9-11-11-11-31 to 8-10-9-28, leaving the secondary timings alone. This was done at their rated vDIMM of 1.60 V.
The result was a perfectly stable (well, SuperPi 32M stable) RAM at much tighter timings. This is actually quite relevant to the RAM market. You see, there are only three DDR3-2133 4 x 8 GB kits on Newegg, and all of them are G.Skill. Talk about cornering the market.
When you drop speed down to DDR3-1866, there is one 4 x 8 GB kit with timings similar to these on the Egg. It is made by Corsair and rated at DDR3-1866 / 9-10-9-27 / 1.5 V. Those timings aren’t as tight as these but its voltage rating is lower. The kicker is that kit costs $200 more than this G.Skill kit. This is no cheap RAM, but that price difference is insane!
Next up, we’ll crank some voltage (up to 1.70 vDIMM) and see how far this kit can be pushed. First up is the most difficult benchmarkers use – SuperPi 32M.
Truthfully, when G.Skill mentioned it this kit didn’t strike me as easy to overclock…until I tried it. Boy was I wrong. A rather conservative 1.70 vDIMM led to SuperPi 32M stable DDR3-2304! Impressive.
It didn’t stop there though, with SuperPi 1M completing as high as DDR3-2322.
The maximum frequency I could validate was DDR3-2332. I say that with emphasis because I actually grabbed .cvf files (CPUz validation files) as high as DDR3-2342 but the validations were corrupted when I tried to submit them.
Surprised doesn’t begin to describe my reaction to this. The IMC on this CPU is quite weak and can’t even make it to DDR3-2400 with RAM rated for it, so loosening the timings was a moot point – this kit took the IMC to its max at their default timings!
Final Thoughts & Conclusion
In a word: Impressive. Not only did G.Skill send a 32 GB kit of RAM rated at blistering-for-the-density speeds, it even overclocked past what any normal person would need. SuperPi 32M stable at DDR3-2304 is stellar! Sure, you have to raise the voltage to get there, but that’s what you’re supposed to do to overclock, isn’t it?
For stock stability, it doesn’t get more rock solid. The performance falls right where it should be and no one should be surprised there.
The price is even impressive. You’re going to pay a pretty penny for a 32 GB kit of RAM rated at DDR3-2133. In this case, that price is $399.99 shipped from Newegg. G.Skill has the DDR3-2133 market cornered, both by the fact that they are the only manufacturer there and that their price is very competitive.
The Corsair kit mentioned earlier (rated at DDR3-1866 / 9-10-9-27) costs $599.99. The biggest difference is going to be its overclocking ability. That kit is rated at 1.5 V, while this one is rated at 1.6 V. That might hint at improved clocking and it may well out-clock this kit; we don’t have one to test. 0.1 V isn’t something I’d be rushing to spend an extra $200 on though!
The long and short of it is this: If you need 32 GB of very fast RAM to plow through whatever it is you do that requires that much RAM, this is a solid kit that does what it says it will do and then some. Its overclocking ability is just icing on the cake. Adding a lifetime warranty, the total package makes for a kit that is easily Overclockers Approved.