970 memory allocation issue revisited

Discussion in 'Videocards - NVIDIA GeForce' started by alanm, Jan 23, 2015.

Thread Status:
Not open for further replies.
  1. Fox2232

    Fox2232 Ancient Guru

    Messages:
    11,809
    Likes Received:
    3,366
    GPU:
    6900XT+AW@240Hz
    It does not cause considerable slowdowns in comparison to fps you get while getting close to 3.5GB utilization.
    But if entire memory got same access condition (each bank accessed via override), your games would have run around 30-50fps at best.

    That is Why I joke, that those who complain so much should e allowed to use entire memory at all times even before card utilizes 3.5GB, so they can randomly load their small 400MB into slow region and wonder why it has 45fps instead of 144fps.
    But question is about real bandwidth as 256bit bus in 980 should perform same as 256bit bus in 970 running same clocks.
    Unfortunately 970 has 224bit bus effectively. Which can be seen as false advertising.
    And if not then expect to see GPUs marketed as 512bit bus where you have 16x32 (512) bits going into memory, but controller itself would be only 256bit having effective speed of 256bit.
     
  2. Fox2232

    Fox2232 Ancient Guru

    Messages:
    11,809
    Likes Received:
    3,366
    GPU:
    6900XT+AW@240Hz
    No, 780Ti would be unaffected:
    If you have mod utilizing 5GB of vram, using 1.5GB per frame you get this:
    - 780Ti fills vram with 3GB of data, uses 1.5GB required per frame at full speed
    - - Once you need data not present in vram, they get loaded from ram and you again access them with full speed
    - 970 fills vram with 4GB of data, uses 1.5GB required per frame, but speed will vary based on place where data are in vram
    - - Once you need data not present in vram, they get loaded from ram and you again face situation where at worst case load 512MB from slow partition.

    You are better off with only 3.5GB accessed.
     
  3. alanm

    alanm Ancient Guru

    Messages:
    10,670
    Likes Received:
    2,766
    GPU:
    Asus 2080 Dual OC
    Aha! So would Nvidia would likely do some driver optimizations for games like this, to block off the .5gb section?
     
  4. WoenK

    WoenK Active Member

    Messages:
    53
    Likes Received:
    0
    GPU:
    Vega 64
    What do the Bits have to do with the GHz?
    if the access frequency is 7 GHz it is almost 4 times faster. And because it is on the board it does not have to go through the PCIE.

    or do you also think that a 2 GHz 32 bit CPU is as fast as a 64 bit CPU with 1 GHz ?
    yeah, right, the moon is made of cheese :D

    As for those claiming that it does not have a 256 BIT controller....none of the 8 memory controllers are switched off, if they would have switched of on they would not be able to access those 0.5 GB.
    http://anandtech.com/show/8935/geforce-gtx-970-correcting-the-specs-exploring-memory-allocation/2

    geez....please all buy a 980, we AMD fanbois do not want your ignorance :D
     

  5. Fox2232

    Fox2232 Ancient Guru

    Messages:
    11,809
    Likes Received:
    3,366
    GPU:
    6900XT+AW@240Hz
    They did already, that is why people see in so many cases only 3.5GB used. Once your fps drops below let's say 35fps, enabling last 512MB will have minimal impact on performance.
    But if your game was running stable 120fps, and then you moved your resource to that 512MB, you would go down to 40fps.
     
  6. Raider0001

    Raider0001 Master Guru

    Messages:
    470
    Likes Received:
    35
    GPU:
    PowerColor RX 6800
    if the access frequency is 7 GHz its 0 times faster because 32bit is 4 times slower than your system memory which is 128 bit when in dual channel
     
  7. Lane

    Lane Ancient Guru

    Messages:
    6,361
    Likes Received:
    3
    GPU:
    2x HD7970 - EK Waterblock
    In reality it is dependant of the Nvidia "optimizations" mode.. so yes it could be an optimization from the driver specific to a game / situation.

     
    Last edited: Jan 27, 2015
  8. goodCode

    goodCode New Member

    Messages:
    7
    Likes Received:
    0
    GPU:
    gtx970 OC
    lol,

    1. Nvidia explained, that 0.5GB is used only when it's needed. So it menas, that usually the cart is working in mode 196GB/s=224bi and 3.5 GB.
    2. All memory parst must have the same memory bandwidth. Example:
    stage a) the card is using only 1 memory segment (3.0 and 0.0).
    stage b), the card must add more 1 GB memory.

    Please calculate the time for operation b?

    time formula: amount/speed.

    theoretical time:1 GB/ 224GB/s= 1/224 s
    practical time: max(0.5GB/196GB/s , 0.5GB/28Gs) = 1/56 s
    using "MAX" formula, because operation is finished, when all 1 GB is passed to the memory.

    so better result is theoretical 1/224 s vs 1/56 s (almost 4 times).

    so is this card with 224GB/s bandwidth? In my calculations - absolutely not!
    because 224GB/s bandwidth must be kept in all situations.

    In my calculations this card's max bandwidth is 196 GB/s.

    Please correct me if I am wrong.
     
    Last edited: Jan 27, 2015
  9. trandoanhung199

    trandoanhung199 Member

    Messages:
    25
    Likes Received:
    0
    GPU:
    Asus EN8500GT
    Well, there goes my plan to eventually SLI it.

    I'll probably just stick with single card, given how many problems there are with SLI/CFX.
     
  10. alanm

    alanm Ancient Guru

    Messages:
    10,670
    Likes Received:
    2,766
    GPU:
    Asus 2080 Dual OC
    And it seems its handled well in most games, even those that use a lot of vram, ie, SoM. But maybe Nvidia need to do more 'specific optimizations' for rare cases like Skyrim, where they may have the game code but not taking into account silly massive texture mods (4096x4096 or above). After all, I believe the main complaints for the 970 vram tanking is based on skyrim. No other games (and I believe Hilbert has tried without replicating the issue). Of course I could make FC4 crawl, but it would anyway with the unrealistic settings I used regardless of the vram issue.
     

  11. WoenK

    WoenK Active Member

    Messages:
    53
    Likes Received:
    0
    GPU:
    Vega 64
    Than maybe endulge us how it is possible to have on a PCIE-E 3.0 slot with 4Ghz and 128/130 bit only transferring 985 MB/s when using a x1 card ? Card has the same 128/130 bit and 4 Ghz.

    Its because of the lanes and non of those "lanes" where cut on the 970, they have disabled some shaders and a cache. Meaning one memory controller has to share one chache with another one, the other 6 have all each their own cache.
     
  12. morbias

    morbias Don TazeMeBro Staff Member

    Messages:
    13,445
    Likes Received:
    37
    GPU:
    -
    Did Nvidia mess up the specs? Yes they did. However, I think people are not seeing what's right in front of their noses; and that's the fact the 970 performs better than it would have if Nvidia had continued what they were doing with previous generations.

    Take a look at the diagram provided by Nvidia engineering:

    [​IMG]

    In the past, the cut-down flagship card would have had two of the memory controllers and double the level 2 cache and ROP segments cut. In other words, if the 970 was to follow previous generations it would only have 3GB VRAM, 1,512kb L2 cache and 48 ROPs on a 192-bit memory interface.

    So essentially the 970 is punching above its weight when it comes to AA and filtering performance and is closer to the GTX 980 than it really should be. Couple that to the fact it is placed at a lower price point than previous cut down flagship cards and it's a bargain.

    There are arguments that it would be faster if it was restricted to 3.5GB but that doesn't hold up either. In that scenario it would go straight to paging the system after 3.5GB, and that would be slower than accessing the remaining 0.5GB at its reduced speed.

    TLDR; the card is closer to the flagship than previous gen and costs way less.
    I think that's something to consider before going on the geforce forums and shouting about entitlement.
     
    Last edited: Jan 27, 2015
  13. War child

    War child Master Guru

    Messages:
    331
    Likes Received:
    4
    GPU:
    Gainward 980Ti SLI
    Im also hearing from users that even though the ROPs are less, GPU Z is still reporting the ROP count as 64.

    Anyone else see this? (I dont own a 970 myself)
     
  14. fantaskarsef

    fantaskarsef Ancient Guru

    Messages:
    12,545
    Likes Received:
    4,905
    GPU:
    2080Ti @h2o
    Been mentioned before (even in this thread I think). GPU Z's data is handfed into an online database, rather than reading out the hardware's specs. That's why it reports something else than is present.
     
  15. Im2bad

    Im2bad Master Guru

    Messages:
    791
    Likes Received:
    0
    GPU:
    3080 Gaming X Trio
    Yes, because that info isn't queried from the card.

    The GPU-Z binary itself has the information on the cards, iirc.
     

  16. Fox2232

    Fox2232 Ancient Guru

    Messages:
    11,809
    Likes Received:
    3,366
    GPU:
    6900XT+AW@240Hz
    Actually 780Ti vs 780 does not look like cutting parts resulted in loss of vram or dramatic drop in performance for part of it.

    Properly cut down 970 would not have 192bit interface and 3GB, but 224bit and 3.5GB what is what it has till last part is enabled, but then it becomes quite complicated in matter of performance to say if it is 192bit + 32bit (shared for 2x32bit with override). Or if it is still 224bit bus accessing vram.

    But as stated many times ago, you are right that performance is what matters and while last 512MB has impact in way it is made to work, it is very small one.

    [mod edit] oops, pressed edit instead of quote, my bad!
     
    Last edited by a moderator: Jan 27, 2015
  17. Razoola

    Razoola Member Guru

    Messages:
    125
    Likes Received:
    2
    GPU:
    GTX980 (MSI gaming)
    I visited the shop today where I purchased my 970. They are aware the situation and are currently talking with their suppliers about what to do for people who want a refund (they know they have to give one). As you can imagine the shop owners that are sitting on a lot of gtx970 stock are not happy, like the one I spoke to today.

    Raz
     
  18. morbias

    morbias Don TazeMeBro Staff Member

    Messages:
    13,445
    Likes Received:
    37
    GPU:
    -
    Well the GTX 780 was the flagship at its release, 770 was the cut down version and the 780 ti came later once the yields were good enough.

    Limiting the card to 3.5GB would give worse performance than having the extra 0.5GB of slower onboard memory as it would be paging the system instead, or at least that's what I gather from the description of how the memory system works.

    [EDIT] oh wait, you're right; I forgot the 770 used the silicon from the 680. Nevertheless, the spec difference was huge between the two high end mainstream cards (770 and 780)
     
    Last edited: Jan 27, 2015
  19. khanmein

    khanmein Ancient Guru

    Messages:
    1,647
    Likes Received:
    72
    GPU:
    EVGA GTX 1070 SC
    r u sure can refund? i also wish my country able to accept refund.
     
  20. Loophole35

    Loophole35 Ancient Guru

    Messages:
    9,793
    Likes Received:
    1,148
    GPU:
    EVGA 1080ti SC
    I got what you were saying. Kinda like the 580 and 570
     
Thread Status:
Not open for further replies.

Share This Page