1. This site uses cookies. By continuing to use this site, you are agreeing to our use of cookies. Learn More.

Graphics Nvidia’s GTX970 has a rather serious memory allocation bug

Discussion in 'Hardware' started by lancer778544, 23 Jan 2015.

  1. Corky42

    Corky42 Where's walle?

    Joined:
    30 Oct 2012
    Posts:
    9,648
    Likes Received:
    388
    TBH you are taking me a little out of my depth :geek:
    So i hope someone else with a better understanding than me can correct any mistakes.

    From my understanding from Windows Vista onwards drivers got moved into user mode kernel space, it's why Vista was so flaky because vendors were learning how to write user mode drivers versus the old way of kernel mode drivers.

    Basically drivers could no longer communicate directly with the hardware without going through the HAL for stability reasons.
     
  2. Gareth Halfacree

    Gareth Halfacree WIIGII! Lover of bit-tech Administrator Super Moderator Moderator

    Joined:
    4 Dec 2007
    Posts:
    17,130
    Likes Received:
    6,719
    I want to live in Theory; everything always works there...
    My point was that, as I understand it, there isn't a caching algorithm - at least, not one written by Nvidia. It would be the game engine that pushes textures it might need soon into VRAM, not the graphics card driver. Now, there may be additional optimisations Nvidia can do - "if we've accessed this texture more than X times in the last Y seconds, move it from 'slow' VRAM into 'fast' VRAM," - but the actual "fill whatever VRAM we haven't already used up with upcoming textures" should be a game-engine thing as the driver has absolutely no idea what textures the engine will need in the near future. I think(!)
    Which lends more weight to my theory that a 3.5GB 'fast'/512MB 'slow' VRAM 970 is more powerful than a 3.5GB 'fast' 970 - because look how crappy TurboCache was due to having to rely on system RAM rather than VRAM.
    Note the phrasing: 'can basically be considered to be' is not the same as 'is,' otherwise it wouldn't need to say 'can basically be considered to be.' A car can basically be considered to be a horseless carriage, but it isn't a horseless carriage 'cos a horseless carriage wouldn't go very far unless it was pointing down a steep hill with the brake off. The HAL sits between the driver and the hardware.

    Quick digression, 'cos I'm procrastinating again (and, if nothing else, this thread has been great for that!): originally, programmers had to write drivers that talked to each bit of hardware individually. You could make a graphics card that would make a game look pretty, but that game had to support that exact graphics card. When IBM came up with the 'Personal Computer,' it came with a BIOS and the programmers were given strict instructions: you don't talk to the hardware any more, you talk to the BIOS which talks to the hardware on your behalf. Boom: the first proper hardware abstraction layer. That was great when IBM was the only company making the BIOS, but the 'IBM Compatible' clones soon followed - and what worked on IBM's BIOS didn't necessarily work on Company B's BIOS. Awkward.

    So, along comes Microsoft with Windows. One of its main selling points is a hardware abstraction layer in software: you don't talk to the hardware, you don't talk to the BIOS, you talk to the Windows HAL which handles all the rest of it for you. Compatibility++ with minimal effort. This would directly influence the development of the Windows Games SDK, which would become DirectX.

    So, yes, the HAL is like a driver, but it isn't a driver. It also isn't written by Nvidia; it's written by Microsoft (with input from Nvidia, AMD, Intel, and any other company you can think to name.)
     
  3. Corky42

    Corky42 Where's walle?

    Joined:
    30 Oct 2012
    Posts:
    9,648
    Likes Received:
    388
    I'm happy to be corrected, but i think Nvidia has in fact written a caching algorithm for how it handles the last 512MB of slower vRAM, unless i got hold of the wrong end of the stick and misunderstood what caching you were referring to.

    Probably in a similar fashion as a CPU has a caching algorithm for how it handles it's caches.
     
  4. loftie

    loftie Multimodder

    Joined:
    14 Feb 2009
    Posts:
    3,173
    Likes Received:
    262
  5. Kronos

    Kronos Multimodder

    Joined:
    6 Nov 2009
    Posts:
    13,495
    Likes Received:
    618
    I have pretty much decided to pay the extra and get a 980 and hope she does not find out. There is far to much uncertainty as to whether this issue is going to prove a headache for Nvidia. But with the various forums I have looked at and given that most of the posts are technically way past my pay grade the consensus seems to be don't buy a 970.
     
  6. Corky42

    Corky42 Where's walle?

    Joined:
    30 Oct 2012
    Posts:
    9,648
    Likes Received:
    388
    PCPER done some more testing on frame times.
    http://www.pcper.com/reviews/Graphics-Cards/Frame-Rating-Looking-GTX-970-Memory-Performance
    If his testing has shown that some performance settings are more likely to cause stutter in the sub-25 FPS range, my guess is that in SLI it would show similar results around the sub 50 FPS range.

    For SLI i would probably agree that the 970 may not be ideal at extreme settings, but as a single card solution most people won't see any problems (unless you enjoy playing at sub 25 FPS)
     
  7. Kronos

    Kronos Multimodder

    Joined:
    6 Nov 2009
    Posts:
    13,495
    Likes Received:
    618
    I think the key word here is most. I know it would bother me, the slightest glitch in a game I would immediately think that it was this issue was to blame and what a fool I was to buy the 970 ETC ETC. So, send her out onto the streets and get the cash for a 980 and hopefully have less stress in the long run.
     
  8. heir flick

    heir flick Minimodder

    Joined:
    2 Feb 2007
    Posts:
    1,049
    Likes Received:
    14
    just read on overclockers, evga have agreed returns on there 970's
     
  9. wyx087

    wyx087 Homeworld 3 is happening!!

    Joined:
    15 Aug 2007
    Posts:
    11,994
    Likes Received:
    713
    I said "just over 30", sorry I didn't remember the exact average FPS 7 years ago :sigh:. There is also memory bus limitations and total memory limitations between GTS and GTX.

    Whereas the 970 is sold as only have SM clusters disabled. Even with ROP disabled, the ratio should still be proportional.


    Anandtech article suggested there is in fact a caching algorithm used.
    The article didn't point out where this algorithm sits. As pointed out earlier, this could be optimised in the firmware or even memory controller (best case) or rely on driver optimization for each game (my worst fear).

    Thank you for the explanation. It's great to learn something new every day.

    So if HAL reports VRAM. Does that mean there's no way to change the amount of VRAM the game sees unless the graphics card are always limited to 3.5GB, thus HAL only senses 3.5GB?


    My knowledge resides in hardware architecture design, I've little knowledge in drivers and game optimizations. Hence I'd rather have a solid hardware to begin with.
     
  10. Kronos

    Kronos Multimodder

    Joined:
    6 Nov 2009
    Posts:
    13,495
    Likes Received:
    618
    Here is the statement.

    HI there

    We have EVGA with us today and they are first to provide an update on the GTX970.

    If you are completely dissatisfied with your EVGA GTX970 then EVGA will 100% support any return of your product through OCUK.

    EVGA do however encourage you to contact them directly before returning your card, as options to step up to GTX980 will be made available but only those in direct contact.

    Please contact supportEU@evga.com if you have any concerns with your EVGA GTX970 product.
     
  11. Corky42

    Corky42 Where's walle?

    Joined:
    30 Oct 2012
    Posts:
    9,648
    Likes Received:
    388
    AFAIK the HAL would query the registers of the hardware, so it just reports what the hardware presents to it.
     
  12. Gareth Halfacree

    Gareth Halfacree WIIGII! Lover of bit-tech Administrator Super Moderator Moderator

    Joined:
    4 Dec 2007
    Posts:
    17,130
    Likes Received:
    6,719
    Then we've both learned something today - I hadn't seen that. Thanks for the link!
     
  13. andrew8200m

    andrew8200m Multimodder

    Joined:
    4 May 2009
    Posts:
    2,507
    Likes Received:
    264
    http://www.techpowerup.com/209409/p...s-being-returned-over-memory-controversy.html

    I do find this rather amusing...

    It doesn't matter on the spec. It could be 1 of everything, ROPs, shaders #whatever. The launch reviews had X performance, the cards still have X performance. Information about the way the 970 is technically a little broken aside as if it wasn't brought to light and there would be no issues.

    Mountain
    Mole hill

    As im not in retail or anything like that anymore I can say the following.

    Customers are now and have always been idiots. end of.
     
  14. wyx087

    wyx087 Homeworld 3 is happening!!

    Joined:
    15 Aug 2007
    Posts:
    11,994
    Likes Received:
    713
    Interesting, I'll give EVGA a buzz, I wonder if I can step up to 980 for around £100.

    Worth a try, can't hurt in asking.
     
  15. Parge

    Parge the worst Super Moderator

    Joined:
    16 Jul 2010
    Posts:
    13,022
    Likes Received:
    618
    It'll be £200 won't it - the cost of the difference between the two cards.
     
  16. Kronos

    Kronos Multimodder

    Joined:
    6 Nov 2009
    Posts:
    13,495
    Likes Received:
    618
    Absolutely worth a try. Don't ask don't get as they say.
     
  17. rollo

    rollo Modder

    Joined:
    16 May 2008
    Posts:
    7,887
    Likes Received:
    131
    Not really sure why anyone cares. The card is dirt cheap yet blows away anything accept the 980 above it which costs £200 more at last check. Are people bored with lack of other GPU News or just hoping for discounts.

    If it had 1 rop 1 cuda core and 1mb of memory but delivered the frame rates it does would it make a difference? Outside of some hardcore elite that must have specs at set levels.

    People are making crazy assumptions about memory usage in 4k without real benchmarks to back it up. Theres not been a game released yet that has been designed for 4k. Till it happens performance at that resolution varies so much on a game by game and setting by setting scenario that how much memory your gpu has or hasnt is kinda irelivent.

    GPU memory is only really needed for High AA settings, Game Devs are not going to start putting 6gb of texture files to the gpu buffer. As the Consoles they are designing around do not have that level of buffer to play with. (PS4 has 4.5gb, Confirmed on ps4 website by infamous second son developer)
     
  18. Harlequin

    Harlequin Modder

    Joined:
    4 Jun 2004
    Posts:
    7,131
    Likes Received:
    194
    you appear to have missed the last few days - this isn't some little matter - dying light at 1080p is filling a 970`s ram.
     
  19. N17 dizzi

    N17 dizzi Multimodder

    Joined:
    23 Mar 2011
    Posts:
    3,234
    Likes Received:
    356
    If you plan to go 4k and SLI I would drop the 970, as usage more regularly goes into the slower partition at 4k in my experience with it. I wish I had 2 980s to compare side by side.

    After this debacle I hear people are planning to sue the makes of Neverending Story.
     
    Last edited: 30 Jan 2015
  20. Harlequin

    Harlequin Modder

    Joined:
    4 Jun 2004
    Posts:
    7,131
    Likes Received:
    194
    Nvidia have withdrawn the statement it will use drievrs to fix the hardware issues
     

Share This Page