The Kits

UPDATE: Curious how DDR4-2133, 2666, and 3200 perform in the hottest games? Then jump to our Feb. 2016 article on DDR4 Gaming!

Introduction

The impact of system memory on gaming performance is a topic lots of gamers are interested in, but finding in-depth analysis can be hard. Back in mid-2013, we published our analysis of how RAM quantity affects gaming performance in our article Does Memory Matter? 4GB, 8GB, and 16GB in Gaming. That article has gone on to be one of the most popular articles on The Tech Buyer's Guru, to this day! So it’s with great pleasure that we bring you the next chapter in our RAM benchmarking analysis, a deep dive into the impact of RAM speed on gaming performance. Now, we need to get one thing out of the way first – this is most definitely not an overclocking article. If you want to squeeze the very highest performance out of your hardware through hours (or days) of tweaking, there are plenty of other websites that have demonstrated how you can do that for individual products. Instead, we are taking one very highly-spec’d RAM kit, the TridentX from G.Skill, and running it through its paces at five of the most popular RAM frequency/timing combinations on the market today. Each of these combinations is at or below the published specs of the TridentX kit – the point here is to show you what the RAM specifications do for you, the gamer, rather than to show you how high this individual kit can go in the hands of a dedicated overclocker (hint - it can't go much higher!).

Each of the RAM combinations we tested is available from various manufacturers at various prices, and in our opinion, RAM is a commodity – if you want DDR3-1600 with timings of 9-9-9-24, most kits labeled as such are essentially identical, and will perform exactly the same if run at their published rating. There probably is some variation at the margins with less established or more overclocking-friendly manufacturers, but if you’re using RAM from any of the manufacturers we recommend on this site, you’re in good hands. That means Corsair, Crucial, Patriot, G.Skill, Mushkin, Kingston, and several others from time to time. In our own personal and benchmarking systems, we are currently running a combination of Corsair, G.Skill, and Crucial kits, along with some awesome (but sadly discontinued) low-profile Samsung RAM. We wouldn’t hesitate to use them all again...even in mixed sets. It’s no longer true, and probably never was, that you’ll get better stability or performance by using matched sticks of RAM – remember RAM is a commodity, and if it’s the same size running at the same spec, it’s basically the same product. In fact, many RAM kits literally use the same synchronous dynamic random-access memory (SDRAM) chips, since there are actually very few worldwide manufacturers of such chips.

Here are representative RAM kits matching the settings we tested, along with current retail prices:

As mentioned at the outset, we didn't actually test each of these RAM kits, although we have several of them in-house, used in various PC builds. Instead, we took the highest-rated kit, the G.Skill TridentX, and downclocked it to simulate the other kits. Because all RAM operating at the same frequency and timings will work identically regardless of vendor, we consider this an accurate way of depicting performance without needing to physically change RAM for each benchmarking session. Note that RAM speed is determined both by frequency (measured in MHz) and its timings. The most important timings are CAS latency (CL), RAS to CAS delay (tRCD), RAS precharge (tRP), and cycle time (tRAS), which are the four timings included in the list above. Of slightly lesser importance is the command rate (CMD). We set it at 1T, rather than 2T, for all our tests, but 2T offers wider compatibility, especially when running four sticks of RAM.

Increasing frequency almost always requires a loosening (increase) of timings, and sometimes also necessitates an increase in voltage. Striking the right balance is critical, and in our opinion is handled quite well by manufacturers already, hence the various timings and voltages in the kits listed above. We've found that tweaking RAM settings has very little payoff in terms of performance, particularly given the increased chance for system instability, so running at a kit's specified settings (provided in the XMP profile) is typically a good approach. To get a rough approximation of RAM speed, you can divide frequency by the CAS latency, but as we'll see in the benchmarks that follow, this is truly just an approximation.

One note - we hinted at this above, but we'll say it more directly here: do not buy lower-spec'd RAM with the hope of overclocking it to the next tier of RAM. It just isn't going to happen, at least not without big voltage increases, which we generally frown upon. At stock volts, our DDR3-1600 kits could not hit the DDR3-1866 settings we used, and likewise, our DDR3-1866 kits could not hit DDR3-2133. So just buy the RAM kit that most appeals to you (and your wallet). The RAM vendors have done the overclocking for you, so don't look a gift horse in the mouth, just take it for a ride!

The benchmarking system we used has the following specs:

Now, how about those games? Well, we wanted to provide a broad cross-section of genres, and we also wanted to test relatively new, more demanding games. At the same time, we had to stay away from brand-new games, because it’s extremely challenging to benchmark software that's in the middle of constant updates. So, here are the games/applications we tested, listed in the order in which they appear in this article:

  • 3DMark
  • Grid 2
  • Metro: Last Light
  • Thief
  • Crysis 3
  • Battlefield 4

OK, let's have a look at 3DMark to get some baseline data....

3DMark

3dmark

3DMark receives a lot of criticism from gamers, because it isn’t actually a game. That’s true, but it’s still a very useful utility, and we always use it to set a baseline for both CPU and video card performance. Not surprisingly, the Graphics Score is essentially identical in all five test cases, as it should be. Any difference you see is simple testing variance – that much variation would be present even when testing the exact same system five times in a row. More interesting for purposes of our exploration here is the Physics Score, which is a pure CPU platform benchmark – the video card is used only to display the image as rendered by the CPU. And as such, it’s a nearly perfect test of theoretical CPU power. From one setting to the next, the difference in scores is sometimes barely measureable, but from lowest to highest, we see a delta of 2.1%, which we consider significant. It's not a big difference, but it's there. Assuming that the physics test employed here is more sensitive to RAM speed than actual game rendering, we'd be surprised if the five game benchmarks we have in store for you show more than that 2.1% delta.

Let's see if that theory holds true, shall we?

Next page