Games becoming more bandwidth sensitive?

Discussion in 'Games, Gaming & Game-demos' started by Carfax, Apr 23, 2020.

  1. Carfax

    Carfax Ancient Guru

    Messages:
    3,088
    Likes Received:
    584
    GPU:
    Zotac 4090 Extreme
    When was the last time any of you looked at memory bandwidth benchmarks for games?

    The last time I looked at benchmarks on bandwidth sensitivity in games was years ago. During the Nehalem days with triple channel memory, and especially when the x79 platform came out for the Sandy Bridge core based 3900 series CPUs that used quad channel interfaces. In those days, games were decidedly much more different than they are today as the PS3 and Xbox 360 were alive and well. Single threaded 32 bit games using high level APIs like DX9, DX10 and DX11 were the norm. These games could barely register a performance difference between one or two threads, let alone single, dual, triple or quad channel memory.

    Unlike today where practically every game is 64 bit from the ground up, can use at least 4 threads (even with DX11), and many of the more advanced game engines like Frostbite 3, AnvilNext, Id Tech 7, UE4 etcetera now make excellent use of low level API optimizations and can make use of high core count CPUs with ease.

    So things have obviously changed a great deal since the advent of the PS4 and Xbox One, and for the better.

    Anyway, I was having a discussion on another forum about the possible impact that DDR5 will have on CPUs and one guy mentioned that games tend to be much more sensitive to latency than they are to bandwidth. My first instinct was to agree with him, as that was my experience from several years ago as well. But then I checked a recent video on YouTube that tested a 6800K (six cores) with 4x4GB and 2x8GB of DDR4 3000 CL16 in BF5 and AC Odyssey with a 1080Ti running at 1080p high settings.



    The results astonished me. The performance difference between them was HUGE. On average, I was seeing 10-30% difference in average framerates, which is akin to a GPU upgrade. But what really floored me, was the difference in 1% lows. There were several times when there was over a 100% difference in favor of the quad channel rig, which was amazing! :cool:

    Though to be honest, should we really be surprised? Frostbite 3 can use up to 8 threads, and AnvilNext can use up to 10 threads. Games have been becoming more and more parallel over time. This trend will increase with the next gen consoles as well.

    Greater parallelism should in theory result in more emphasis on memory bandwidth, and less on latency. This is a good thing, because memory bandwidth is easier to improve compared to memory latency.
     
    Dragam1337 and jbscotchman like this.
  2. coc

    coc Maha Guru

    Messages:
    1,054
    Likes Received:
    492
    GPU:
    3060Ti
    this is new to me. interesting.
     
  3. jbscotchman

    jbscotchman Ancient Guru

    Messages:
    5,872
    Likes Received:
    4,758
    GPU:
    MSI 1660 Ti Ventus
    Good post, I've also noticed this trend increasing. 2x8 gig has been the sweet spot for PC gaming for a while, but that will definitely change with the new consoles. I'll be adding another 16gb soon.
     
  4. Carfax

    Carfax Ancient Guru

    Messages:
    3,088
    Likes Received:
    584
    GPU:
    Zotac 4090 Extreme
    Well when DDR5 comes out, it won't be much of an issue because bandwidth is going to go WAY up. You will be able to get over 100GB of memory bandwidth with just dual channel memory, unlike with today where you need quad channel and better to hit those numbers.
     

  5. Carfax

    Carfax Ancient Guru

    Messages:
    3,088
    Likes Received:
    584
    GPU:
    Zotac 4090 Extreme
    I think you're referring to memory capacity, which is different. Memory capacity is the amount of RAM that you have (which is also important), but my post was about memory bandwidth, which is the amount of data that can be accessed (read, write and copy) by RAM. All current mainstream CPUs, both Intel and AMD use dual channel memory. This can result in being bandwidth limited, especially for high count CPUs like the 3900x and 3950x.

    Intel has quad channel interface for its HEDT CPUs, and hex channel for their Xeon lineup. I think AMD uses octa channel memory for Epyc CPUs.

    Anyway, games have historically not been very responsive to memory bandwidth and more so to latency, but this appears to be changing as games and 3D engines become more parallel by design.
     
    Dragam1337 likes this.
  6. gerardfraser

    gerardfraser Ancient Guru

    Messages:
    3,343
    Likes Received:
    764
    GPU:
    R9 290 Crossfire
    I am not convinced by the Quad channel being that much better ,otherwise when testing quad channel on reviews, these results would have been shouted out loud.
    I would have bought Intel Quad channel setup for sure 100%.I think the video in the first post is off,something is not set correct but I do not mind being wrong and if this is true I would only build machines with Quad channel or AMD 8 channel.I would agree that some games may have a postive performance impact on dual vs quad channel.

    As far as dual channel goes ,if one knows how to setup ram there would be no real difference in PC gaming unlike what everyone thinks about latency and bandwidth.

    BF5 and GTAV 1920x1080 DDR4 2933Mhz vs 3866Mhz no difference in gaming.

    RDR2 and Sleeping Dogs 1920x1080 DDR4 2933Mhz vs 3866Mhz no difference in gaming.
     
  7. JonasBeckman

    JonasBeckman Ancient Guru

    Messages:
    17,563
    Likes Received:
    2,960
    GPU:
    XFX 7900XTX M'310
    Fallout 4 scaled really well with memory speeds, not sure if the channels were tested as frequently but there was quite a boost from DDR3 stock to the higher speeds like 1800 let alone what modern DDR4 could do.
    Watch_Dogs and other open world games really seem to benefit too though what these go through once on a SSD now that's a transformation ha ha.

    Ghost Recon Breakpoint loading in less than a minute for the main menu and then a few seconds to the game world itself from that, might not be a immediate framerate changer but still among the most impressive SSD improvements for gaming I've compared.
    Guessing for memory there's a few other titles too really benefiting and even more so now with the AMD processors really liking this infinity clock high-speed and low latency performance around 3600 - 3800 speeds or just above if it's synced and stable.

    Suppose it will also be important once the next-gen games come around and start to really utilize higher speed I/O as a default and how bottlenecked that could get but that's not a immediate change though it will be interesting to see how this gets utilized as HDD's are no longer as much of a factor for performance and issues with data loading.


    EDIT: Really going to be nice when the hardware is working together and the limiting factor for startup and loading times is just the software default of really lengthy really slow unskippable opening logos and a feature film length equally unskippable end-of-game credits list. :p

    Less "forced walk" masked loading times is going to be a design changer too with getting rid of a majority of these narrow passages, infinity elevators and pipes and whatever else. :)

    ...Never getting rid of that "Press some button to actually start." main menu holder though I suppose but oh well there will still be other improvements. :D


    EDIT: There's that Fallout 4 memory scaling benchmark.
    From Techspot. :)

    [​IMG]


    Current AMD and Intel processors and improvements, DDR4 and I expect it to still benefit nicely even if it already sees good gains here for 2400 DDR3 on the Intel system and a decent bit on the earlier AMD hardware too.

    If AMD does triple or quad channel on the non-Ripper boards that might also see some benefits but we'll see when that happens if there's any major gains to be had from supporting these outside of the Threadripper boards and the complications this involves and additional increase in motherboard prices and such for the end users.

    Though I suppose a TX40 3990 DDR4 3800 test on Fallout 4 or other memory sensitive current titles would also work heh wouldn't surprise me if some of these could still see impressive scaling.
    (Sorta, weird how much of a bottleneck that actually is on some of these considering the locked console hardware for the XBox One and PS4 and what that would have as a baseline hardware target but it's a thing.)
     
    Last edited: Apr 24, 2020
    AsiJu and Dragam1337 like this.
  8. Dragam1337

    Dragam1337 Ancient Guru

    Messages:
    5,326
    Likes Received:
    3,399
    GPU:
    RTX 4090 Gaming OC
    You are gpu limited, so ofc you wont see any difference with ram speeds -_-'

    For you to see any difference with ram speeds, you need to be cpu limited - not gpu....
     
    Last edited: Apr 24, 2020
  9. Dragam1337

    Dragam1337 Ancient Guru

    Messages:
    5,326
    Likes Received:
    3,399
    GPU:
    RTX 4090 Gaming OC
    I never tested duel vs quad channel either, but ram speed does indeed affect fallout 4 alot - witcher 3 even more so. Tested it a long time ago in this post :

    https://imgur.com/gallery/ivovA
     
  10. TheDeeGee

    TheDeeGee Ancient Guru

    Messages:
    8,479
    Likes Received:
    2,481
    GPU:
    PNY 4070 Ti XLR8
    Oh wow, thanks for the heads up!

    Planning to go Ryzen (3700X probably) later this year when B550 boards come out, so this is really valuable information!
     
    Last edited: Apr 24, 2020

  11. JonasBeckman

    JonasBeckman Ancient Guru

    Messages:
    17,563
    Likes Received:
    2,960
    GPU:
    XFX 7900XTX M'310
    For the Ryzen processors the 1:1 speed of the infinity fabric and the ram speed the resulting performance and latency improvements make it well worth to try and go for 3200 - 3600 DDR4 kits, lower latency is nice but the price tends to ramp up and some kits that also overclock very well or chips tend to have a price premium to where it might not be entirely worth it.

    B550 should have the newest and latest AGESA and other bios improvements too whereas for the X570's and such the bios flashing to the latest available hopefully also bundling the latest AMD AGESA updates should be among the first things done for compatibility and general fixes and improvements, beyond that the AMD power profile and chipset drivers and it's good to go preferably with Windows 10 Redstone 5 or the current Vanadium (20H1) builds for underlying improvements like the thread scheduler and kernel. :)

    RAM and infinity fabric scales up to around 1800 - 1900 before it goes into 1:2 mode and requiring some manual tweaking, think the max official supported was just above 3800 though kits like 4000+ work but requires some bios editing which for Ryzen and newer AMD boards this isn't as straightforward or easy although far as overclocking goes loading the XMP is generally enough and other tweaks will either be detrimental or require careful testing and comparison since you are balancing it for each CCX and per core and such factoring in the CPU cores overall load and how this scales so it's often best left to default even if there are gains to be had but it's more tricky to set up and also to avoid possible degradation from pushing too much voltage and AMD's own scaling to temperature and other factors which also changed in newer AGESA updates and bios patches. :)


    It'll have a bit more latency than comparable Intel systems and it's dual channel instead of tri or quad but it's usually really fast and performance shows a good bump over the default 2400 specs of DDR4 though I wouldn't spend a fortune on specific memory kits unless you're doing some enthusiast build trying to max it out.
    (Two 16 GB kits here, worked really well even pre bios flashing - A bit of a oops on my end there. - which was lucky because it can be a bit sensitive.)


    EDIT: As to performance well I knew a few titles benefit from fast RAM like Fallout 4 but it's quite something seeing that it wasn't just a CPU bottleneck particularly getting the HDD speed changed into a SSD even if it's not the latest as PCI-E 4.0 NVME 1.4 kits only started showing up when I built this system but it'll hold nicely until it either starts degrading or I can build the new system whenever that time comes. :)

    Think there's a video comparison for Watch_Dogs 2 and I assume Far Cry and possibly GTA would also really benefit, others might still be more CPU or GPU limiting but RAM could still make a bit of a difference.


    Need to see if I can find that video.

    EDIT: Nope but there's a few other ones. :)


    Such as this one using a Intel 9900k processor and a few different titles showing some nice gains.

    Witcher 3 does show some good gains and it's a city environment for that particular test so the CPU has something to work with so it might scale even more once that's less of a limiting factor depending on how things go and how well the game engine scales with additional cores and such and how much the CPU would be held back if it can't scale across more than 4 - 6 cores or so.
     
    Last edited: Apr 24, 2020
    Carfax and Noisiv like this.
  12. AsiJu

    AsiJu Ancient Guru

    Messages:
    8,233
    Likes Received:
    2,952
    GPU:
    MSI RX6800XT G.XT.
    @JonasBeckman
    Ryzen 3000 scale IFCLK and MCLK 1 : 1 up to DDR4 3733.

    Going past that a 1 : 2 ratio kicks in but you can manually set IF clock higher even then to close the gap.
    However using faster than 3600 memory doesn't really seem to give any practical benefit.

    Between 3000 ... 3600 is the sweet spot and the lower latency you have the better.
    So 3000 CL14 memory can actually be better than say 3600 CL18.

    Though I don't know what would happen if you had say 4500+ speed memory, would the increased bandwidth start making a difference again.

    I remember seeing a chart for mem speed scaling for Intel 7700 or 8700.
    From 3200 up to 4000 you got pretty much diminishing returns but going past 4000 started making a difference again, interestingly.
    The scaling was as good if not better than between 2400 and 3000.

    Grain of salt ofc but interesting anyway.
     
    JonasBeckman likes this.
  13. vestibule

    vestibule Maha Guru

    Messages:
    1,162
    Likes Received:
    489
    GPU:
    Radeon RX6600XT
    I remember playing AVP2 on dial up. No problems and it was mad and great fun but expensive.
    Oh, that kind of bandwidth. Hey ho. I will get my coat. :)
     
    AsiJu likes this.
  14. Amaze

    Amaze Ancient Guru

    Messages:
    3,987
    Likes Received:
    1,135
    GPU:
    3060 ti 2075/7800
    Am4 won't support quad band so a moot point for me :oops:
     
  15. AsiJu

    AsiJu Ancient Guru

    Messages:
    8,233
    Likes Received:
    2,952
    GPU:
    MSI RX6800XT G.XT.
    Btw possible explanation why 4 DIMMs can be better than 2 DIMMs on a dual channel configuration when using single rank DIMMs (from about 0:50 to 4 min mark, followed by benchmarks):



    Indeed minimum framerates might get a nice boost.
     
    Last edited: Apr 29, 2020
    JonasBeckman, Carfax and Noisiv like this.

  16. jiminycricket

    jiminycricket Master Guru

    Messages:
    203
    Likes Received:
    4
    GPU:
    GTX 1080
    Yes, dual rank modules can perform similarly to single rank modules clocked a few hundred MHz higher. At the same time though, it can be counterbalanced by dual ranks per channel being more stressful for the IMC, so it may not be able to clock as high as single rank per channel.
     
    JonasBeckman and Noisiv like this.
  17. Carfax

    Carfax Ancient Guru

    Messages:
    3,088
    Likes Received:
    584
    GPU:
    Zotac 4090 Extreme
    Yes, I've been meaning to update this thread but I kept forgetting. It appears memory interleaving is responsible for the performance gain you get with dual rank vs single rank memory. Having dual rank allows the memory access pattern to be more parallel and uniform which reduces latency and increases throughput.

    Memory interleaving used to be a setting in the BIOS that you could enable or disable a long time ago, but now I guess it fell under the radar because motherboard manufacturers aren't allowing people to disable it. And why should they, when disabling it reduces performance.
     
  18. Astyanax

    Astyanax Ancient Guru

    Messages:
    14,950
    Likes Received:
    6,120
    GPU:
    GTX 1080ti
    memory interleaving in that case was ganged vs unganged channel interleaving.
     
  19. teleguy

    teleguy Maha Guru

    Messages:
    1,363
    Likes Received:
    221
    GPU:
    GTX 1070/Vega 56
    Noisiv likes this.

Share This Page