Slashdot Mirror

← Back to Stories (view on slashdot.org)

NVIDIA Responds To GTX 970 Memory Bug

Posted by timothy on from the can't-remeber-why-you'upset dept.

Vigile writes Over the past week or so, owners of the GeForce GTX 970 have found several instances where the GPU was unable or unwilling to address memory capacities over 3.5GB despite having 4GB of on-board frame buffer. Specific benchmarks were written to demonstrate the issue and users even found ways to configure games to utilize more than 3.5GB of memory using DSR and high levels of MSAA. While the GTX 980 can access 4GB of its memory, the GTX 970 appeared to be less likely to do so and would see a dramatic performance hit when it did. NVIDIA responded today saying that the GTX 970 has "fewer crossbar resources to the memory system" as a result of disabled groups of cores called SMMs. NVIDIA states that "to optimally manage memory traffic in this configuration, we segment graphics memory into a 3.5GB section and a 0.5GB section" and that the GPU has "higher priority" to the larger pool. The question that remains is should this affect gamers' view of the GTX 970? If performance metrics already take the different memory configuration into account, then I don't see the GTX 970 declining in popularity.

5 of 145 comments (clear)

  1. This reminds me... by YuppieScum · · Score: 1, Informative

    ...about 10 years ago, we had some performance issues related to the NVidia cards included with a large batch of HP workstation-class PCs purchased by the investment bank I worked for.

    We had sufficient clout to drag in the local NVidia reps to answer our questions... after a bit of a grilling, and discussion of the reference numbers on the various bits of silicon soldered to the cards, they admitted that the total RAM claimed to be on these cards (64Mb IIRC) was actually half on the card and half from system RAM, and required a specific version of the driver to make work. Without that driver, you had half the video RAM you had expected... and paid for.

    Since then, none of the unscrupulously weaselly behaviour of this company has surprised me, and they've not had a penny of my money nor of any company I worked for since.

    --
    This sig left unintentionally blank.
    1. Re:This reminds me... by pushing-robot · · Score: 5, Informative

      You're describing 'TurboCache' (a marketing name if ever there was one).

      It wasn't a secret, it was only on very low end cards, and ATI was already doing the same with 'HyperMemory'. Intel, for their part, was exclusively using system RAM at the time (and largely still is).

      So what graphics *have* you been buying for the last decade?

      --
      How can I believe you when you tell me what I don't want to hear?
  2. Re:False Advertising by Gravis+Zero · · Score: 3, Informative

    I'm just going to come out and say that to advertise the card with 4GB, but then disable any amount of it, is false advertising.

    i agree. however in this case, all 4 Gigabytes are accessible, they simply aren't accessible at the same speed. the final 500MB is "slow" to access but it's still there and you can still access it.

    --
    Anons need not reply. Questions end with a question mark.
  3. Re:False Advertising by Solandri · · Score: 3, Informative

    I'm just going to come out and say that to advertise the card with 4GB, but then disable any amount of it, is false advertising. Sure, most games can't actually hit 4GB since most games are still brain-dead 32-bit applications that can't access more than 4GB of any memory.

    But this is a sign of things to come. Where the next generation sub-20nm GPU's will be advertised with RAM amounts and supposed to have 2-3X the processing power, but part of the GPU will be competely unusable because the operating system or software being used isn't 64-bit aware.

    VRAM has nothing to do with system RAM. VRAM is special memory used by the dGPU, and only the dGPU, for storing framebuffers, textures, models, and other data needed to draw a 3D scene. It's faster than system RAM (GDDR5 is typical, vs DDR3 for regular RAM), and positioning it closer to the GPU reduces latency due to the speed of light (which travels only 10 cm in a single 3 GHz cycle). So the 32- or 64-bitness of the OS and apps has nothing to do with the video card's ability to access 4GB or more of VRAM.

    In particular, the 970 GTX has a 256-bit memory bus. The speed constraint of having to retrieve data from VRAM one 32-bit (float) or 64-bit (double) "chunk" at a time became a bottleneck long before the inability to address that VRAM as a flat memory space. So mid- and high-end video cards are designed to retrieve multiple "chunks" of data from VRAM simultaneously. You have to drop all the way down to the GT 730 before you get to video cards using a 64-bit memory bus.

  4. Re:Hey! I've been gypped! by wierd_w · · Score: 3, Informative

    Silly person!

    Use UMBPCI instead of EMM386, and use CTMOUSE for the mouse driver.

    (assuming your modern system still knows how to play right in real mode anyway. Many modern chipsets have problems with ISA style DMAs, which makes using the hardware UMBs free with UMBPCI can have unpredictable results. For such systems, you are stuck with EMM386 doing protected mode memory reassignments, and gobbling down a big chunk of conventional. Blech.)

    Really, there are much better memory managers that came about since the DOS days (FreeDOS is still a living project for devices that simply must run DOS. Industrial vinyl cutters and the like come to mind), and you can reasonably get over 568k conventional free with little hassle.