Announcement

Collapse
No announcement yet.

The NVIDIA vs. Open-Source Nouveau Linux Driver Benchmarks For Summer 2018

Collapse
X
  • Filter
  • Time
  • Show
Clear All
new posts

  • #21
    Originally posted by cybertraveler View Post
    For whatever reason, you're not acknowledging the economic principle I explained. Your last comment is a reflection of that. If NVIDIA are charging a premium for compute/workstation hardware that is no more capable than their consumer hardware (with artificial restrictions removed), then they likely either hold a near-monopoly position or have superior tech.
    Your economic principle is correct, but as I tried to explain you are ignoring other restrictions that apply here (and in other fields too), which prevent such from making sense in this case.

    R&D costs to make a new GPU generation are a single thing, you can't just develop gaming GPUs or businness/computing GPUs cards separately, so the end product has to be technically able to do both, to an extent (number of stream processors and so on of course limits the power of a card, but the features are mostly the same across the board, the overall architecture is the same).

    The point here is that you can't split evenly the costs on each card based on its actual raw power. It would rise the price too much for gaming cards, they won't sell.

    You can't sell all such cards at gaming card pricing either, as that won't cover all your R&D costs.

    And they can't just decide to not sell gaming cards at all and focus on the other market as that would again increase their costs and decrease the sales of the businness and compute cards, which is again bad.

    This is why product design isn't a job for sissies.

    So what they do is treat it like a service. You pay some money to get a device that provides you some level of 3D acceleration in games, which can be seen as a "service".
    The fact that the hardware itself might be actually able to do anything else is a technicality, an artifact of what the technology in this field can provide, you did not pay for the development of those features when you bought your "gaming device" and should therefore have no access to them.

    The price they are selling the consumer cards at is a price that includes only the 3D performance and gaming features. NVIDIA is adding a premium on that because they can, and some competition could surely help on reducing the premium. But this won't decrease the prices of businness or compute cards, as that's another market segment, paying for more features. Even basic Quadro that have raw 3D performance equivalent to 300$ cards can cost thousands of dollars.

    People have trouble understanding that it actually works like this with most physical products, especially with electronics you're not paying that much the hardware itself (the physical hardware costs relatively little in materials as it's a mass-produced component), but R&D costs to get there with a product that can do many different things.

    So you hear people saying "but why they made this board and then populate only half of it!?! They paid for the device already right?"... no they don't. They designed the product so they could install features as needed.
    This way they could sell the product at a cheap enough price for you while keeping features in more expensive devices so they can sell them at higher prices, while both devices are basically coming off the same R&D budget, and neither would allow them to get enough money to cover it all.

    For example (an example of someone that is less likely to be evil like NVIDIA) the Raptor Engineering company that made the PowerPC motherboards for consumer market has now provided a second board that slashes away a whole lot of components and features that quite frankly don't cost shit hardware-wise.
    It's a bunch of connectors and dumb passive components that won't cost more than 400$ per board if bought by a consumer, much less if it is bought in bulk at manufacturing scales.

    But they are offering that same board at around half the price of the full-feature one, around a 1500$ price reduction when they saved only like 400$ of components. Must be madness right?

    If they designed the board from scratch it would have costed more or less the same as the other instead, 2500$ or more.

    This stuff is the real job of product designers and marketing, deciding how to split the costs of R&D over products by making tiers of features and pricing so they can maximize the sales and therefore best cover the costs and actually make and maximize the profit.

    their competitors will be able to undercut them by selling equivalent compute/workstation hardare to the same market at a lower price (a price much closer to the consumer market GPU hardware which you are arguing is capable of compute/workstation performance when not artificially restricted by software).
    Given what I said above, their competitors have the same limitations, so they don't.

    They can compete within the market segment, but not from a different segment.
    You won't see a FirePro selling for what its 3D gaming value would be, that would mean AMD is selling it at a loss.

    Also: some gamers do want and use GPU VM passthrough features.
    That's a niche and you know it. GPU/VGA/pcie passthrough is an arcane art that only a few linux and ESXi users master, and is completely unknown outside of those circles.

    That said you can still freely use KVM instead of Xen, it works fine with NVIDIA cards (one of the reasons I said that GPU firmware alone can't really stop passthrough). Recommended to use the passthrough setups relying on the card's UEFI bios (the card must have it. All cards that have less than 2 years can be assume to have that, for older ones you need to check for it by googling), as that's the best in terms of performance and ease to set up.

    You are fixated on this mid-range card thing. IT WAS AN EXAMPLE.
    An example that makes no sense, because the premises of it ignore the actual limitations of the field, which I explained above.

    After reading that, of course I am going to believe you're talking about making midrange cards (and other non-high-end cards) run worse.
    I will concede that I was not 100% clear on that so I'll let you get out free on this.

    I'd just like to point out that "more expensive ones" could have also meant the businness cards. You can play games fine on a Quadro or a FirePro. It's just kind of wasteful to drop 1-2k $ for a card that performs more like a 400-600$ gaming card and has other features you don't need.

    Comment


    • #22
      Thanks for the thoughtful reply starshipeleven. I am not completely convinced by your arguments though. I have my own theory for why even physically identical hardware could be sold for a much greater price to a compute or workstation audience than it would a consumer audience. It relates to what you said about selling a service not a physical product. My theory is that the compute & workstation customers are paying a premium for these things:
      • Specific application support.
      • Compatibility promises for the specific applications they use their hardware with. This may mean that NVIDIA promptly produces patches to fix issues with those applications and to remain compatible with new versions of those applications.
      • Promises to "support" the specific hardware functions they are using (eg the graphics card passthrough feature you mentioned). A gamer may have a consumer card which is physically capable of those functions, but NVIDIA may not advertise those functions to the gamer and may make no promises that those functions will work for any specific purpose. The compute / workstation customer however can rely on that support and even get legal recourse in the event that NVIDIA did not deliver that support; this is because NVIDIA would have specifically sold a product with advertisements claiming it did something when it did not: this is a form of fraud and theft by deception.
      • SLAs
      So, if NVIDIA doesn't have a monopoly position and/or more advanced tech than their compeition, they can still charge extra for compute / workstation cards that may be physically identical to some consumer/gaming cards, because of the above. NVIDIA would not need to hold back an Open Source driver implementation in this scenario, because even if an Open Source driver was created, the Open Source driver would not provide the user of it with those benefits listed above (SLAs, developers focused on specific app support, compatibility promises and support for specific card features).

      As such I am currently of the belief that in recent history NVIDIA has had a near-monopoly on the gaming market or simply had vastly superior tech which its competitors have struggled to rival.
      Last edited by cybertraveler; 06-16-2018, 12:34 PM.

      Comment


      • #23
        Originally posted by cybertraveler View Post
        My theory is that the compute & workstation customers are paying a premium for these things:
        • Specific application support.
        • Compatibility promises for the specific applications they use their hardware with. This may mean that NVIDIA promptly produces patches to fix issues with those applications and to remain compatible with new versions of those applications.
        • Promises to "support" the specific hardware functions they are using (eg the graphics card passthrough feature you mentioned). A gamer may have a consumer card which is physically capable of those functions, but NVIDIA may not advertise those functions to the gamer and may make no promises that those functions will work for any specific purpose. The compute / workstation customer however can rely on that support and even get legal recourse in the event that NVIDIA did not deliver that support; this is because NVIDIA would have specifically sold a product with advertisements claiming it did something when it did not: this is a form of fraud and theft by deception.
        • SLAs
        Please note, also AMD's FirePro cards do the same.

        Also please note that "specific applications support" means they support some features used by CAD or modeling applications or whatever, like some OpenGL compat profiles (it's some compatibility mode for old stuff) with a decent performance. It's not just certification.

        Only difference is that AMD can't support CUDA because of obvious reasons.

        And afaik both do not just "not advertise" the functions they don't want gaming cards to have, they disable them in the driver or in the hardware. Afaik amd does not cripple computing performance in their gaming cards though.

        AMD traditionally played the good guy on CPU where they are leaving ECC "enabled but not certified" on consumer hardware, but not on GPUs.

        NVIDIA would not need to hold back an Open Source driver implementation in this scenario, because even if an Open Source driver was created, the Open Source driver would not provide the user of it with those benefits listed above (SLAs, developers focused on specific app support, compatibility promises and support for specific card features).
        That's a nice "truth", it would be very sad if actual facts would poke holes in it.

        They cripple their gaming card computing performance (double precision performance, and with later cards also something else too), this isn't a myth, it's an accepted fact that their gaming cards have much less of the computing performance of what is a more or less equivalent card sold for business use. The card hardware is mostly the same, the crippling is done by the driver. An opensource driver could not enforce this.
        There are official statements too https://devtalk.nvidia.com/default/t...88540/#3288540

        Yes, full-speed double precision performance is a feature we reserve for our professional customers. Consumer applications have little use for double precision, so this does not really affect GeForce users. Having differentiated features and pricing is actually fairer for all. Given the option of enabling all professional features on GeForce and having gamers pay for them, or disabling them on GeForce and offering a more compelling price, we feel the latter is the better choice.

        And tests https://arrayfire.com/explaining-fp6...mance-on-gpus/
        Last edited by starshipeleven; 06-16-2018, 01:33 PM.

        Comment


        • #24
          « but for those who care just about gaming on Linux and not the driver license, the official NVIDIA 396 Linux driver continues performing great with OpenGL/Vulkan/OpenCL/CUDA.»

          Unfortunately, this isn't even an option if you own a laptop. You just have no way to use your hardware then.
          i3+gt640m works way better than i7+gtx1050.
          No way i'm buying an « Nvidia (under)powered » laptop ever again.

          Comment


          • #25
            My experience with Linux and Nvidia (as a user who doesn't do much gaming) has been one of disappointment and despair, especially with the introduction of the 390 drivers. The most important thing for me is a smooth and responsive desktop, and in this regard, the proprietary Nvidia drivers just don't deliver. Sure, they may be able to pump a gazillion pixels to the screen every second - but something they can't do is implement smooth scrolling and a stutter free desktop experience. Even without re-clocking and running at lower clock speeds, the Nouveau driver delivers a much "smoother" desktop experience. I've spent days confirming this - install the Nvidia 396.45 driver, enable wobbly windows, and use Firefox. Scrolling in Firefox exhibits micro-stutters, wobbly windows stutters, sound via HDMI doesn't always work. It's a terrible end-user experience. Uninstall the proprietary Nvidia drivers and use Nouveau, and there's almost (unfortunately "almost") no stutter at all when performing the same operations (and as an added bonus, sounds via HDMI works flawlessly).

            I wouldn't mind Nvidia's pathological obsession to guard their "secret sauce" if they delivered a proprietary Linux driver that worked awesomely, but they haven't (and every indication is they never will).

            I've just ordered an AMD card, will see how the AMDGPU/AMDGPU-PRO drivers are with the Linux desktop.

            EDIT: Just one common use case where the Nvidia drivers perform much worse than the Nouveau drivers:

            A GTX 1070 with 6GB, two monitors attached: a 1080p and a 1440p. Machine with 32GB of RAM running Kubuntu 18.04. I normally open two instances of Firefox - one on the 1080p plays YouTube videos, while the other Firefox instance sits on the 1440p monitor and is used for general browsing etc... Scrolling in Firefox in this scenario is atrocious - it makes it look like there's no GPU driver installed at all. Very, very, very choppy.

            Repeat the same with the Nouveau driver, and things work awesomely (everything is ultra-smooth and without stutter).
            Last edited by ScottDeagan; 08-04-2018, 11:02 PM.

            Comment


            • #26
              Originally posted by starshipeleven View Post
              AMD does not seem to care about such grave risks as they have low market share to begin with, and their main priority now is selling stuff, not necessarily enact a master plan for the future like this. If it turns out they start selling a ton of consumer GPUs and less pro ones they can scale accordingly, they don't have so much to lose as NVIDIA that has basically cornered the market in the high end.
              We do care (we have to), but keeping consumer drivers closed-source is not the only way to deal with the risk. Between optional closed-source userspace drivers and fused-off HW features at chip level there are enough other ways to keep sufficient differentiation between brands.

              John Carmack summed it up pretty well over a decade ago:

              The workstation vendors do stupid driver tricks to make CDRS go faster, while consumer vendors do stupid driver tricks to make Q3 go faster.
              Only the apps have changed
              Last edited by bridgman; 08-07-2018, 02:59 PM.

              Comment


              • #27
                Man the 900 and 1000 series performance is depressing.

                Comment

                Working...
                X