Sign in with
Sign up | Sign in

Battlefield 3, Frame By Frame

Memory Scaling, AMD's Trinity APUs, And Game Performance
By

Long frame times are most jarring to me when there's a lot of on-screen movement. While slowing down usually helps mask this phenomenon somewhat, that's not really a viable workaround in first-person shooters and racing games.

We've established that it's difficult to record evidence of this phenomenon in multi-card configurations. But Fraps does make this possible in single-GPU systems. We're using it today to record performance in Battlefield 3.

It's difficult to generalize, but many folks can tolerate a 20 FPS minimum. So, we set an upper limit of 50 ms per frame to assure reasonable fluidity. Beyond that, adding time per frame can be a much more intrusive distraction.

The sad fact is that even with an average of 50 FPS (shown on the previous page), our fastest memory configuration can't reliably keep the A10-5800K's on-board graphics processor under 50 ms per frame.

Of course, maximum rendering times get worse as resolution increases. Memory latency could be an issue, but even pricey low-latency kits are barely better than the DDR3-1600 CAS 7 config we tested, or this setup's DDR3-2133 CAS 9 arrangement.

Display all 93 comments.
This thread is closed for comments
Top Comments
  • 22 Hide
    Onus , February 4, 2013 3:36 AM
    Quote:
    And there were a few games where the faster memory was required simply to break us into our minimally-accepted 40 FPS average.

    Right there. An APU is not a top-tier gamer, so incremental improvement really matters. I could not care less about going from 60FPS to 80FPS, but 30FPS to 40FPS, the same relative improvement, is a really big deal.
  • 21 Hide
    esrever , February 4, 2013 4:17 AM
    Can't wait to see what DDR4 can do for APUs.
  • 16 Hide
    s3anister , February 4, 2013 5:43 AM
    merikafyeahConsidering how DDR3-2400 is only a tiny fraction better than DDR3-2133, it's safe to assume memory stops being the bottleneck around that point. DDR4 will not noticeably improve performance or even power consumption as memory consumes almost negligible amounts of electricity to begin with.It's back to looking at better GPUs and CPUs for better performance.Bpttleneck hierarchy has always been GPU>CPU>RAM.The CPU has always been more reliant on the RAM than the GPU but an APU is basically a GPU+CPU in one, so RAM is more important, but as we've seen, only up to DDR3-2133. After that diminishing returns skyrocket.

    Oh you are so close yet so far to knowing what you're talking about...

    You would be well served to learn about the Von Neumann architecture and more precisely the Von Neumann bottleneck.

    http://en.wikipedia.org/wiki/Von_Neumann_architecture

    The biggest bottleneck in any architecture is shared communication between all components, data throughput is crucial to all parts of a system; yet beyond that latency of components in relation to bandwidth is the real Achilles heel of the computer and THAT is why CPUs have L1/L2/L3 cache so that they can have ultra low latency memory that is usually around 1.5/5/7.5ns respectively. When you have that low latency combined with a bandwidth of what is 76000/44000/22000 MB/s compared to normal DDR3-1600 on sandy/ivy bridge you have a system that appears to not be bottlenecked by RAM. As for a Trinity AMD system, the reason why one sees such massive gains when going up in RAM speeds from DDR-3 1600 to 2133 is because the GPU can't get away with the tiny amount of storage that is the L1/L2 cache, it has to have a large interface of 512MB-3GB to crunch the massive amount of parallel data and therefore is limited by the aggregate throughput of the system's memory. Hypothetically, if you were to continue to increase the data rate of the system's memory you would see performance gains up until the point where the GPU's instructions units can no longer make use of the available interface.

    Having said all that, until DDR-4 is out we can't say for certain that it will not have a huge impact on both AMD and Intel systems. This is because if DDR-4 manages to lower latency or greatly increase bandwidth you will see gains, especially if DDR-4 is able to achieve both lower latency and higher bandwidth at the same time. Oh and, to correct your first inaccuracy, DDR-4 will be lower power than what is currently available so it will use less electricity than DDR3-2400 therefore providing more performance per Watt of energy used.
Other Comments
  • 4 Hide
    hmp_goose , February 4, 2013 3:22 AM
    Neat stuff, but how sensitive is an IGP to timings?
  • 6 Hide
    Wisecracker , February 4, 2013 3:28 AM
    Thanks.

    The question is ... does the performance with higher speed memory continue to scale as the *SIMD Engine Array* is over-clocked.

    Inquiring minds would like to know ...

  • 22 Hide
    Onus , February 4, 2013 3:36 AM
    Quote:
    And there were a few games where the faster memory was required simply to break us into our minimally-accepted 40 FPS average.

    Right there. An APU is not a top-tier gamer, so incremental improvement really matters. I could not care less about going from 60FPS to 80FPS, but 30FPS to 40FPS, the same relative improvement, is a really big deal.
  • 21 Hide
    esrever , February 4, 2013 4:17 AM
    Can't wait to see what DDR4 can do for APUs.
  • -5 Hide
    slomo4sho , February 4, 2013 4:34 AM
    slomo4shoI would really like to see some 5760x1080 benchmarks.


    Individuals who would use faster memory for gaming are likely to want to push their mid/high range card to the limits, do you plan on doing a similar piece for AMD CPUs as you did in the Intel article "Does Memory Performance Bottleneck Your Games?"

    Also, I would like to see a Nvidia card at play as well. Maybe a 650 Ti or 660 Ti? In addition, it wold be nice to see the memory scaling difference between AMD and Nvidia GPUs in a single review.

    Thanks.
  • 9 Hide
    jubas , February 4, 2013 4:39 AM
    slomo4shoAlso, I would like to see a Nvidia card at play as well. Maybe a 650 Ti or 660 Ti? In addition, it wold be nice to see the memory scaling difference between AMD and Nvidia GPUs in a single review.

    I didn't know that nVidia made APU's?
    The more you know... /rollseyes/
  • 8 Hide
    de5_Roy , February 4, 2013 4:45 AM
    quite promising performance from trinity.
    still, 15 gb/s out of ddr3 2400 ram is just sad. i expect amd to improve in the next gen apus. the igpus deserve the extra memory bandwidth.
    i wonder how cpu overclocking (along with igpu and ram oc) affect the games like skyrim, starcraft and f1. those seemed more memory sensitive.
  • -2 Hide
    austenwhd , February 4, 2013 5:11 AM
    When will they integrate a gpu to FX module, that will be blast.
  • 16 Hide
    s3anister , February 4, 2013 5:43 AM
    merikafyeahConsidering how DDR3-2400 is only a tiny fraction better than DDR3-2133, it's safe to assume memory stops being the bottleneck around that point. DDR4 will not noticeably improve performance or even power consumption as memory consumes almost negligible amounts of electricity to begin with.It's back to looking at better GPUs and CPUs for better performance.Bpttleneck hierarchy has always been GPU>CPU>RAM.The CPU has always been more reliant on the RAM than the GPU but an APU is basically a GPU+CPU in one, so RAM is more important, but as we've seen, only up to DDR3-2133. After that diminishing returns skyrocket.

    Oh you are so close yet so far to knowing what you're talking about...

    You would be well served to learn about the Von Neumann architecture and more precisely the Von Neumann bottleneck.

    http://en.wikipedia.org/wiki/Von_Neumann_architecture

    The biggest bottleneck in any architecture is shared communication between all components, data throughput is crucial to all parts of a system; yet beyond that latency of components in relation to bandwidth is the real Achilles heel of the computer and THAT is why CPUs have L1/L2/L3 cache so that they can have ultra low latency memory that is usually around 1.5/5/7.5ns respectively. When you have that low latency combined with a bandwidth of what is 76000/44000/22000 MB/s compared to normal DDR3-1600 on sandy/ivy bridge you have a system that appears to not be bottlenecked by RAM. As for a Trinity AMD system, the reason why one sees such massive gains when going up in RAM speeds from DDR-3 1600 to 2133 is because the GPU can't get away with the tiny amount of storage that is the L1/L2 cache, it has to have a large interface of 512MB-3GB to crunch the massive amount of parallel data and therefore is limited by the aggregate throughput of the system's memory. Hypothetically, if you were to continue to increase the data rate of the system's memory you would see performance gains up until the point where the GPU's instructions units can no longer make use of the available interface.

    Having said all that, until DDR-4 is out we can't say for certain that it will not have a huge impact on both AMD and Intel systems. This is because if DDR-4 manages to lower latency or greatly increase bandwidth you will see gains, especially if DDR-4 is able to achieve both lower latency and higher bandwidth at the same time. Oh and, to correct your first inaccuracy, DDR-4 will be lower power than what is currently available so it will use less electricity than DDR3-2400 therefore providing more performance per Watt of energy used.
  • 2 Hide
    csbeer , February 4, 2013 6:47 AM
    What is the best discreet video card to match with that setup to run in asymetrical xfire and how would that stand up against a similarly priced intel + dgpu setup? I have a feeling what the answer is, but I want to believe AMD! Make me believe!
  • 2 Hide
    m32 , February 4, 2013 6:55 AM
    Not bad, but not impressive. I can't wait till SteamRoller APUs come out. That should be the next big step for iGPUs..... unless Haswell shows something increible.
  • 7 Hide
    Crashman , February 4, 2013 7:35 AM
    hmp_gooseNeat stuff, but how sensitive is an IGP to timings?
    It's in there, check the DDR3-1600 C9 vs DDR3-1600 C7 numbers. Basically, switching from C9 to C7 gives you a performance benefit roughly similar to one speed upgrade.
  • 1 Hide
    Anonymous , February 4, 2013 9:26 AM
    personally I would like to see comparion benchmarks to see how the apu stacks against a dedicated graphics, same proc just with a dedicated graphics card
  • 5 Hide
    Memnarchon , February 4, 2013 9:31 AM
    Actually the growth (in power) of the APUs make them need faster memory bandwith. They are screaming for the DDR4 to become available at the market.
  • 4 Hide
    whyso , February 4, 2013 10:03 AM
    How about hybrid crossfire. That's something that AMD markets a lot and there is few reviews about.
  • 1 Hide
    cypeq , February 4, 2013 10:45 AM
    RAM speed always matters for integrated GPU because it doesn't have it's own dedicated super speed memory.
    RAM Speeds above ddr 1333 does not bottleneck any current CPU in terms of gaming.
    Hardly any game gove above 2 GB or ram used so 4 GB is what you need 95% of time you buy 8 GB because it's cheap.
    And that's all about memory performance.
  • 1 Hide
    Soda-88 , February 4, 2013 11:06 AM
    For SC2 to be playable competitively (at any level really), it needs to run at 60FPS from 6 supply to the end of the game. Spiking between 20 and 40 is too frustrating to play such a high APM game at (some pros like DRG and Sniper, both GSL champions, spike well over 500).
    That's the main reason why you see a lot of streamers stream at the lowest possible settings which makes game look like crap but provides required fluidity to avoid the frustration of looking at the slideshow in which you can't act and proceed to lose the game.
  • 0 Hide
    sanilmahambre , February 4, 2013 11:28 AM
    APU trinity is for a budget gamer and i believe that every part he choose must fit his pockets specially when the difference ratio is so less
Display more comments