Haswell And Richland Memory Scaling: Picking A 16 GB DDR3 Kit

Maximizing Integrated Graphics: Data Rate Or Latency?

With data rates topping 6 GT/s on buses up to 384-bits wide, high-end graphics cards are hardly starved for memory bandwidth. Integrated engines, on the other hand, are still limited to a 128-bit aggregate pathway through two 64-bit channels. And then there are the lower data rates of DDR3 DIMMS versus GDDR5 packages. But throughput isn't the biggest issue those built-in GPUs face. They're necessarily a lot less complex, since they share die space with host processing resources. And then there's the matter of memory latency...

If we look back to when DDR2-533 CAS 3 was the way to go for exceptionally low latency, we saw a few enthusiasts trying to push those same modules down to CAS 2. These days, we see DDR3-1600 CAS 7 as a real possibility, and most tuners are having similar trouble pushing that memory to CAS 6. Even at DDR3-2133, we’re trying to hit CAS 8 when only CAS 9 appears achievable. Lo and behold, when we divide 2133 by eight, 1600 by six and 533.3 by two, we always get 266.6. Divide by two to get the bus frequency and invert the number to get cycle time, and what we’ve really been fighting for the past nine years is a memory turnaround time of 7.5 nanoseconds.

Still unable to reliably break past the 7.5 ns latency barrier, today’s performance search primarily focuses on data rates. We’ll still compare these six contenders to JEDEC-standard DDR3-1600 CAS 11 specs though. Here are each kit's specifications.

Swipe to scroll horizontally
Brand/ModelRated MT/sRated CLRated VoltsPrice
Adata XPG V2 AX3U2400W8G11-DMVDDR3-240011-13-13-351.65 V$184
AMD Gamer Series AG316G2130U2KDDR3-213310-11-11-301.65 V$176
Crucial Ballistix Tactical BLT2KIT8G3D1869DT1TX0DDR3-18669-9-9-271.50 V$153
G.Skill DDR3-1866 C10 F3-14900CL10D-16GBXLDDR3-186610-11-11-301.50 V$135
Mushkin Redline Ridgeback 997121RDDR3-21339-11-11-281.65 V$180
Patriot Viper 3 PV316G240C0KRDDDR3-240010-12-12-311.65 V$195

We chose JEDEC-standard DDR3-1600 as a starting point because we believe that anyone seriously interested in gaming on a platform with integrated graphics shouldn’t settle for less. We found a 16 GB dual-channel pair of those for $110.

Thomas Soderstrom
Thomas Soderstrom is a Senior Staff Editor at Tom's Hardware US. He tests and reviews cases, cooling, memory and motherboards.
  • Someone Somewhere
    What about 8GB? That's still a good option for most gamers.
  • DjEaZy
    ... GPU 4.T.W.!!!
  • Novuake
    Its a shame the AMD memory controller is so lacking, it matters very little for FX series, but here its a real bummer.

    Anyway, I am reasonably sure you could tighten the timings on the Ripjaws considerably, if this is the case with this specific kit, there is NO reason to get the Mushkins...
  • Jarmo
    8GB of fast&expensive memory vs 16GB of the cheap stuff?
  • CommentariesAnd More
    IMO 4GB and 8GB is what most APU Gaming Rigs have. Those looking for an APU for Gaming probably would love results with similar kits. But still we would love these results for like a baseline. Also , the Elite FTW in my next APU build/proposal :D
  • yannigr
    You did it again. Another "let's make Intel looks better than it is" article.

    How can we give an advantage or two to Intel?
    Let's overclock both cpus. So Intel is getting 1GHz extra from it's default frequency while AMD only gets 400MHz. Then we call this fair.
    Let's lower the resolution and graphics. That way the benchmarks will be less gpu intensive and Intel's much better cpu will make things look more even.
    Not to mention the instability problems with AMD and 2400MHz memory, the problem reaching 4.5GHz. We all know that AMD is unstable and hot. Nice job. You are pros in what you are doing. But off course you where just testing memory here. Yeah right. You lowered the resolution even lower than 1366X768 which is the resolution for hdready displays. I guess lowering the resolution and overclocking was necessary so that Intel gpu can reach 30fps in all tests. Nice work.
    Next time try 800X600 and overclock with LN2.
  • 4Ryan6
    Scuttlebutt says that AMD’s A10-6800K reaches 4.5 GHz without much effort, but our sample needed 1.425 V to achieve complete stability. It would easily run 4.4 GHz at 1.30 V, so perhaps 4.5 GHz wasn’t a good target? The problem with 4.4 GHz was that we didn’t want to give the firm a frequency handicap in an article that includes Intel. Our Core i7-4770K was easily running 4.5 GHz at 1.250 V maximum, and we began testing before AMD's sample arrived. And so the test began with the Intel processor at 4.50 GHz, secure in the knowledge that both CPUs would support DDR3-2400 via overclocking. Two of the kits were DDR3-2400-rated. And one of the kits would even run at DDR3-2666.

    Thomas, I know it would have involved more work, but wouldn't the right thing to do have been to reset the overclock target to 4.4ghz?

    You'd have been in safer territory not to have overclocked at all, if the overclock was not the same for each testing.

    Additionally for the record, the very reason some of these overclockers are experiencing such high heat levels are they're overclocking the CPUs memory controller from the very beginning of their overclocking, by attempting to run memory past the design specifications of the CPU right out of the box, that may have been partly the reason you couldn't get past 4.4ghz with acceptable voltage with the A10-6800K.

    These memory manufacturers never claimed their memory could be run at higher multiplier overclock levels as they were only tested, and supposedly guaranteed at the stock CPU speed capabilities at the factory, they never guaranteed they would run stably at 45x or beyond, multiplier ranges in the first place.

    I have yet to see the first memory claim that BrandX not only can run 2400mhz but can run it at a 50x multiplier for Intel or 25x for AMD, to reach 5.0ghz or for that matter 45x or 22.5x for your 4.5ghz target.

    So actually just because you can push the memory to make these tests doesn't mean that running memory speeds way past the CPUs memory controllers design specifications is a safe 24/7 run solution to advise others is seemingly OK, but that is exactly what a lot of users take away from these type of tests.

    Just because someone can, doesn't necessarily mean they should, for the longevity of their hardware, isn't THGs responsibility on a higher level than that?

  • de5_Roy
    woulda liked to see how these kits would perform on a desktop intel hd5000/5100/5200 sku. shame on intel for not making those available on desktop! hd4600 blows!
  • Novuake
    11546494 said:
    woulda liked to see how these kits would perform on a desktop intel hd5000/5100/5200 sku. shame on intel for not making those available on desktop! hd4600 blows!

    I almost think Nvidia and Intel had some nice discussions and deals to keep Iris out of the desktop.
    Just a hunch...
  • Reynod
    Nice work crashman ... good to see they still have you chained to the test bench out the back.

    I hope they are feeding you buddy and letting you out once in a while !!

    Its time they let you loose on another serious overclocking campaign too.

    Chris ... get this man a big can of liquid nitrogen, 2 cartons of coke, a box of mars bars and a mobile hot dog vendor for the day ... time to reward crash.