In the past several years, enthusiasts have seen storage capacity (and performance) explode. Graphics cards are immensely faster. And CPUs operate more efficiently.
System memory doesn't evolve as quickly, though. Just look back to the math I did in Haswell And Richland Memory Scaling: Picking A 16 GB DDR3 Kit. We've been stuck with approximately 8.5 ns turnaround time for more than ten years. Really, single-DIMM capacity is the one variable that keeps getting nudged up.
Today, enthusiasts are pushing 8 GB modules closer to the mainstream. This capacity isn't new. However, technical issues kept it out of the fastest builds up until recently. Increased density often imparts a latency penalty, and those modules had to hit a sub-9 ns sweet spot before the power users out there would call them performance parts. Intel pushed progress along a little with a more robust memory controller in its Haswell-based CPUs, and ratings like DDR3-1600 CAS 7 (8.75 ns), DDR3-1866 CAS 8 (8.57 ns), and DDR3-2133 CAS 9 (8.44 ns) became the benchmark for 16 GB dual-channel kits.
I covered the 16 GB market in that previously-linked round-up. However, the battle between capacity and performance continues for anyone who wants more. As with increased density, adding modules also has a negative effect on overclocking capability.
Recently, I noticed a spate of press releases in my inbox extolling four-DIMM DDR3-3000 kits that never actually showed up for purchase. A majority went to sponsored overclockers, I figured. But they got me thinking: just how far will a quartet of today's top modules go when we load down a Haswell-based system with 32 GB?

| Rated Specifications | ||||
|---|---|---|---|---|
| Data Rate | Timings | Voltage | Warranty | |
| 2 x Adata XPG DDR3 AX3U2800W8G12-DGV | DDR3-2800 (XMP) | 12-14-14-36 | 1.65 Volts | Lifetime |
| Corsair Vengeance Pro CMY32GX3M4A2800C12R | DDR3-2800 (XMP) | 12-14-14-36 | 1.65 Volts | Lifetime |
| G.Skill Ripjaws X F3-2400C11Q-32GXM | DDR3-2400 (XMP) | 11-13-13-31 | 1.65 Volts | Lifetime |
| Kingston HyperX Beast KHX24C11T3K4/32X | DDR3-2400 (XMP) | 11-13-13-31 | 1.65 Volts | Lifetime |
| Patriot Viper 3 PV332G240C1QK | DDR3-2400 (XMP) | 11-13-13-31 | 1.65 Volts | Lifetime |
We only invited companies that could provide us with four-DIMM, 32 GB sets. Adata ended up sending in two of its 16 GB dual-module kits, which we concluded was fair enough, since enthusiasts are certainly within their right to double up on a configuration like that to hit 32 GB. But the company could run into trouble if its modules are programmed to operate with one module per channel. Often, two DIMMs per channel require relaxed latency settings.
| Products |
ADATA XPG DDR3 (2x)
|
G.Skill Ripjaws X
|
Kingston HyperX Beast
|
| Pricing |
|
|
|
So today, Adata competes against the similarly-rated Corsair kit, which is programmed with the looser secondary timings often needed for two modules per channel at high data rates.
Other contenders enter the fray with DDR3-2400-rated sets priced more conservatively. But really, our primary goal is to identify the paramount module set. Any attractive values we find along the way are purely incidental. This story is all about being the best.
Much of the discussion involving these modules relates directly to the test platform we're using, so it's important to detail our configuration ahead of any questions that arise.

Overclockers have found wide variability between different samples of Intel’s Core i7-4770K, yet the CPU's dual-channel memory controller is usually more forgiving of high frequencies compared to the quad-channel controller found in its Core i7-4960X. My particular sample is better than average, reaching 4.6 GHz at 1.25 V core and DRAM data rates exceeding 3000 MT/s in several of my motherboard round-ups. Higher data rates are important in an overclocking test!

Asus’ Z87-Pro is among the elite memory overclocking motherboards in my stash, and it repeatedly proves itself capable across a wider variety of modules compared to my other top choices.

When I talk about Haswell’s thermal transfer problems, and large coolers barely outperforming smaller ones, Thermalright’s MUX-120 is the impressive smaller one I’m alluding to. This heat sink keeps my Core i7-4770K below 70 °C at 4.5 GHz and 1.25 V.

The reason I’m running the CPU at 4.50 GHz is the same as the reason I’m using PowerColor’s overclocked PCS+ AXR9 290X. I’m trying to build a system powerful enough to expose memory bandwidth bottlenecks in ordinary applications. I could try to do that with an ordinary Radeon R9 290X, but why would I want more noise?
| Test System Configuration | |
|---|---|
| CPU | Intel Core i7-4770K (Haswell): 3.50 GHz, 4C/8T Overclocked to 4.50 GHz (45 x 100 MHz) at 1.25 V Core |
| CPU Cooler | Thermalright MUX-120 |
| Motherboard | Asus Z87-Pro: LGA 1150, Intel Z87 Express, Firmware 1707 (12/13/2013) |
| Graphics | PowerColor PCS+ AXR9 290X 4GBD5-PPDHE: 1050 MHz GPU, 4 GB GDDR5-5400 |
| Hard Drives | Samsung 840 Pro MZ-7PD256, 256 GB SSD |
| Sound | Integrated HD Audio |
| Network | Integrated Gigabit Networking |
| Power | Corsair AX860i: ATX12V v2.3, EPS12V, 80 PLUS Platinum |
| Software | |
| OS | Microsoft Windows 8 Pro x64 |
| Graphics | AMD Catalyst 14.3 Beta 1.0 |
| Chipset | Intel INF 9.4.0.1026 |
| Benchmark Configuration | |
| Autodesk 3ds Max 2013 | Version 15.0 x64: Space Flyby Mentalray, 248 Frames, 1440x1080 |
| WinRAR | Version 5.0: THG-Workload (1.3 GB) to RAR, command line switches "winrar a -r -m3" |
| F1 2012 | Steam Version, In-Game Test Test Set 1: High Quality Preset, No AA Test Set 2: Ultra Quality Preset, 8x AA |
| Metro: Last Light | Steam version, Built-In Benchmark, "Frontline" Scene Test Set 1: DX11, Med Quality, 4x AF, Low Blur, No SSAA, No Tesselation, No PhysX Test Set 2: DX11, High Quality, 16x AF, Normal Blur, SSAA, Tesselation Normal, No PhysX |
| SiSoftware Sandra | Version 2014.02.20.10, CPU Test = CPU Arithmetic / Multimedia / Cryptography, Memory Bandwidth Benchmarks |
Although Adata’s two $570 dual-DIMM kits didn’t exactly align with our original 32 GB round-up announcement, again, we eventually decided that combining 16 GB packages was a completely valid way to achieve 32 GB of memory capacity, and we'd get some good information out of the exercise. After all, these kits feature tighter secondary and tertiary timings compared to some of our four-DIMM sets, and we’d really like to see how that might benefit performance and/or limit overclocking.

Detected as DDR3-1333 CAS 9 via SPD, Adata’s AX3U2800W8G12 kits are easily configured to rated DDR3-2800 at 12-14-14-36 timings by using the motherboard’s XMP setting. That overclocking profile also instructs the board to switch from 1.50 V defaults to 1.65 V.

Adata provides a limited lifetime replacement warranty on all DRAM products purchased through an authorized reseller.
Like Adata, Corsair supplied us with DDR3-2800 CAS 12 parts. But unlike its competitor, Corsair’s kit is designed to operate with all four modules installed. The company adds a fan to move air between tightly-packed modules, and loosens tRFC timings by around 15% to assure stability with the extra parts in place.

The included cooling fan is wide enough to cover six slots, though its clips can be slid around to provide CPU cooler clearance. If red doesn't match the rest of your hardware, you can choose between silver or blue anodized replacement trim.

Booting at DDR3-1333 CAS 9 and 1.50 V at SPD defaults, enthusiast-class motherboards should let you pick this kit's rated DDR3-2800 data rate at 12-14-14-36 timings via its 1.65 V XMP overclocking profile.

A lack of retail distribution leaves buyers only with the choice to purchase these at their full $1500 MSRP from Corsair’s store. At least the parts that arrive carry the firm’s non-transferrable lifetime warranty.
We deal with a ton of convoluted naming, so I really appreciate when company employ easily-decipherable nomenclature. But a good name won't be enough to put G.Skill’s DDR3-2400 CAS 11, quad-channel 32 GB kit on top of this speed race. Instead, these parts need to satisfy my quest for impressive overclocking.

At little more than one-quarter the price of Adata’s DDR3-2800 parts, G.Skill’s $310 kit still has some big overclocking shoes to fill. And just when you're tempted to say, "But I only see two modules," we flip the package over to reveal the other half of this four-DIMM set.

We also get some encouragement from the 11-13-13-31 timings of this kit's DDR3-2400 rating, enabled by Intel’s XMP auto-overclocking technology. Prior to enabling that 1.65 V profile, these parts boot at DDR3-1333 CAS 9 and 1.50 V.

Unlike many of its competitors, G.Skill doesn’t impose a long list of exclusions on its lifetime replacement warranty. Only product abuse and/or mishandling will invalidate the coverage.
Kingston’s attempt to give us a decipherable name yields several abbreviations, though it’s not too difficult to figure out that this is a HyperX, DDR3-2400, CAS 11 kit of four modules totaling 32 GB of capacity. Kingston doesn’t yet have retail packaging for these parts, though the brown box is offered by retailers for as little as $318.

That reasonable price gets buyers four 8 GB DDR3-2400 modules with a secondary DDR3-2133 XMP profile and DDR3-1333 SPD for booting at default voltage levels. Both automatic-overclocks offer 11-13-13 timings, though the higher 2400 MT/s data rate requires an increase from 1.60 to 1.65 V. Once again, we’re counting on extra overclocking headroom to make DDR3-2400 parts compete with DDR3-2800 kits.

Like many of its competitors, Kingston officially provides lifetime warranty coverage only to the original purchaser. Though we’ve never heard any complaints, we recommend saving your receipt.
More extensive use of abbreviations means that, unless you've seen the full model name and specifications of Patriot’s Viper 3 32 GB DDR3-2400 CAS 11 quad-module kit, you'll have a tough time figuring it all out at a glance. When it comes time to search for the best price, copy and paste are your friends.

Specifications for this $380 kit are found on the front label, and the part number is found on the back of the package.

While most competing products boot at a lowly DDR3-1333, Patriot's parts include a DDR3-1600 CAS 9 profile as well. That could make them the perfect test components for my own motherboard reviews, since I only use XMP for overclocking. On the other hand, data rates beyond DDR3-2800 could be a tall order for these DDR3-2400-rated parts.

A simple switch to XMP mode through our motherboard's firmware gives us the full 2400 MT/s data rate, along with 11-13-13-31 timings and a bump to 1.65 V. CPU-Z reports 3T, though the motherboard correctly sets a 2T command rate.
Patriot’s limited lifetime warranty excludes purchases made through unauthorized resellers, in addition to the expected abuse and mishandling exclusions.
Both DDR3-2800 samples are rated at 12-14-14-36 timings, and all three DDR3-2400 samples are rated at 11-13-13-3. But both speeds have much more in common when it comes to the lowest stable latencies that provide optimal response times.
| Best Stable Timings | ||||
|---|---|---|---|---|
| DDR3-1600 | DDR3-2133 | DDR3-2666 | DDR3-2800 | |
| 2 x Adata XPG DDR3 AX3U2800W8G12-DGV | 7-8-8-15 | 9-11-10-18 | 12-13-13-24 | 12-14-13-27 |
| Corsair Vengeance Pro CMY32GX3M4A2800C12R | 7-8-7-15 | 9-10-10-18 | 11-13-12-21 | 12-13-13-24 |
| G.Skill Ripjaws X F3-2400C11Q-32GXM | 7-9-8-15 | 10-11-11-18 | N/A | N/A |
| Kingston HyperX Beast KHX24C11T3K4/32X | 7-8-8-15 | 9-11-10-18 | N/A | N/A |
| Patriot Viper 3 PV332G240C1QK | 7-8-7-15 | 10-11-10-18 | 12-13-12-21 | N/A |
Adata's XPG kit required a slightly higher tRAS (the fourth primary timing, above) setting than the other products when I configured it to operate at DDR3-2666. Lower frequencies paved the way for even lower tRAS numbers than the ones in my chart, but without any improvement in the benchmark results.
Corsair’s Vengeance Pro kit was also able to accommodate the lowest primary latencies. That could be important in the search for the absolute-fastest 32 GB kit.

Higher data rates show well in Sandra’s memory latency test, in spite of the absolute higher latency required to get there. That’s partly because latency is counted in cycles, and cycle time is the inverse of frequency.
We were really surprised that G.Skill didn't send in one of its Trident DDR3-2933 quad-channel kits to take on Corsair’s top-market Vengeance Pro. Unfortunately, a top data rate of DDR3-2600 isn’t going to help the firm win in this overclocking competition.

So, why would a company that does sell a DDR3-2933 kit submit one of its DDR3-2400 offerings for today’s round-up? Perhaps because I said there would be a performance comparison, and G.Skill knew something I didn’t about how this platform would deal with higher data rates.

The top three bandwidth numbers are established at DDR3-2400 settings. Data rates beyond DDR3-2400 require longer secondary timings, and the motherboard responds with looser tertiary timings.
Additional testing confirms that memory bandwidth continues to climb with higher data rates applied to locked-in DDR3-2400 timings, and that DDR3-2400 bandwidth drops to 25 GB/s when manually configured to DDR3-2666 timings. "Best Timings" in the chart labels only refer to the use of optimized primary timings discussed on the previous page.
I kept playing with the timings of Corsair’s top-overclocking modules until I hit 30.4 GB/s at DDR3-2800. That’s still a little short of the DDR3-2400 numbers, but it’s a clear indicator that secondary and tertiary timings are an issue for the Sandra Memory Bandwidth test using Asus’ Z87-based motherboard.
Memory performance has always been a bottleneck in F1 2012, particularly at low resolutions using entry-level quality settings. Of course, that'd be unrealistically light for the hardware we're using. Choosing the lowest-acceptable settings for my high-end test platform, frame rates vary from 159 FPS at DDR3-1600 CAS 9 to 177 FPS at DDR3-2400 CAS 11.

That’s not as significant of a difference as we’ve seen from weaker hardware, though 10% is still noticeable. More than likely, several variables including memory led to our previous findings in F1 2012. I'd guess that processor performance plays a big role, too.

Metro: Last Light is nearly stuck at 117 FPS, even though our Core i7-4770K is clocked to 4.5 GHz and graphics is handled by PowerColor’s overclocked Radeon R9 290X. A second card in CrossFire or an even higher CPU frequency might have revealed memory limitations, but most gamers aren't concerned about bottlenecks at 100+ FPS.
3ds Max doesn’t appear to benefit from bandwidth or latency optimizations, though it does appear to penalize poorly-configured settings just slightly.

G.Skill’s optimized DDR3-2400 timings give it a one-second lead in WinRAR. I was able to repeat that on Corsair’s Vengeance Pro kit at DDR3-2800, but only after manually removing the motherboard’s DDR3-2800 penalty.

It'd seem that the board, rather than the processor's memory controller or DRAM, could be responsible for mediocre differentiation. It's also possible that at these high data rates, additional bandwidth simply isn't being utilized in real-world benchmarks.
The best memory kit in today’s round-up must necessarily be the one that runs at the highest data rates and/or supports the tightest timings. In other words, Corsair's Vengeance Pro DDR3-2800 is the only 32 GB memory kit elite enough to win our Tom's Hardware Elite honor.

That recognition does leave me in a bit of a bind though, since G.Skill took the rules of our contest into account (including the motherboard and settings we told everyone we'd use), and correctly determined that DDR3-2400 would be best-suited to the test platform. Most of us would call that working smarter, rather than harder. And the smart choice for owners of Asus' Z87-Pro will want to pair that board up with the right kit. The right kit should qualify for an award, too.

A value-oriented award was always a possibility in this overclocking-oriented story. But one didn't appear likely until after I saw how wise it'd be to pair our platform with a $310 DDR3-2400 kit. Rather than optimizing for all-out data rate or the lowest latencies, G.Skill tuned for motherboard limitations.
And that opens up yet another can of worms. It would seem that Asus has been winning the DRAM overclocking component of my motherboard round-ups with excessively loose timings when its board is set to aggressive data rates. Sure, I could add DDR3-2800 bandwidth testing to those pieces. But Asus might very well respond by optimizing platforms for data rates no higher than DDR3-2800 (rather than its current DDR3-2400 tuning).
So, which memory would I choose? As a motherboard reviewer, I'm inclined to go with the kit that works best for motherboard round-ups. Since many of today's boards sneakily apply CPU overclocking whenever XMP mode is enabled, I run application benchmarks with XMP disabled. A perfect module set for me would begin with an SPD value of DDR3-1600 CAS 9, such as Patriot’s Viper 3 PV332G240C1QK.
But then I need to add overclocking stability tests. Topping out at DDR3-2693, Patriot’s DDR3-2400 kit just doesn’t add much to motherboard evaluations that employ a DDR3-3000-capable CPU. On the other hand, competing samples default to DDR3-1333 CAS 9 settings that are below our recommended minimum for performance evaluation. That means you'll continue to see me using two different sets of memory for my stories: the Patriot Viper 3 for benchmarks and Corsair Vengeance Pro for overclocking. The best place I can think to put G.Skill's value-leading kit is in my own PC.
At $310 for 32 GB of G.Skill’s DDR3-2400, I can’t fault anyone who follows my lead.