Announcement

Collapse
No announcement yet.

NVIDIA 470 Series Driver Looks Like It Will Bring OpenCL 3.0 Support

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

  • #31
    Originally posted by ColdDistance View Post
    A monopoly built on one sentence I hate in computing: "It just works". OK, CUDA is better, I'm not negating that, but CUDA is a monopoly built on a proprietary technology.

    I see in the future thousands of developers crying and complaining because CUDA is a monopoly, a monopoly they fed for too many years.
    That doesn't answer my question. I'm asking what dubious business practices Nvidia did to make CUDA a monopoly. Of the things I know Nvidia did, they built finely-tuned drivers, they created very in-depth documentation, they created easy-to-use libraries for a wide range of programming languages, and they supplied many example source codes for various purposes that people could follow. How the hell is that dubious? That's what you're supposed to do when you make a product and expect people to adopt it.
    CUDA may be closed-source, but Nvidia isn't preventing anyone else from making CUDA drivers. Considering the time and money they dumped into it, they don't owe anybody else a hand. Why would Nvidia deliberately help their competitors use their product?


    Originally posted by kieffer View Post
    * Nvidia's license which forbids gamer cards to be used within datacenter (except for blockchain). Illegal in most countries anyway. Please don't come with thermal/power issues, this is another business, we know how to deal with.
    THAT is a good example of a dubious practice.
    * Claim we want ECC memory (we don't) and after we set-up an experiment to count how often those events occur (we did not find one in a year on many GPUs), they claimed their have _evidences_ but never shared the protocol with us. We are a radiation facility where creating bit-flip is even part of our business !
    lol where were you in all the arguments I've had in the past about the necessity of ECC? Trust me, there are people here who will vehemently disagree with you on this, and suggest ECC is a necessity.
    That's not to say I feel ECC isn't necessary for Quadros, but having it as an option would be nice. In either case, I don't find this to be all that dubious - Nvidia isn't about to sell an expensive product and risk bit flipping compromising their reputation.
    * twist the arm of top manager and sell DGX system when technical staff (I am part of) tells: Not more than 2 GPU per computer.
    Eh... that's kinda dubious, but that's pretty typical behavior of any business' sales team. I'm not saying Nvidia should be doing this, but if you were to get AMD GPUs, I'd be surprised if you didn't face the same sort of thing.
    Why can't we use OpenCL on ARM or PowerPC systems ? why is the driver version so restricted on those platforms ?
    Yeah, that is pretty weird. Though, it's hard to tell if that was a business decision or just a laziness issue.

    Comment


    • #32
      Originally posted by Qaridarium
      what does OpenCL 3.0 mean ???? they do openCL1.2 and claim it is 3.0 because all features are optional...
      this means openCL3 is a complet failure. Nvidia only claims to support 3.0 because they just support 1,2 and claim it is 3.0
      because they can all features are optional means they don'T support it.
      the people try openCL and end up using CUDA instead because OpenCL is complete insanity(all features are optional)...

      the only fix to this is: USE VULKAN COMPUTE....
      Vulkan compute is the only sane way to stop this insanity they (AMD/Nvidia) can't make a distiction between gaming cards and compute cards...

      to pump up the openCL number makes only sense if the features ARE NOT OPTIONAL...

      any sane person should switch from OpenCL to Vulkan right now.
      See these quotes I found in the below article "If a vendor was happy with OpenCL 1.2 but wanted a couple of extra 2.1 features, for example, then to be compliant with the specification they’d need to implement the entire 2.1 core specification; OpenCL 1.x/2.x had no mechanism for partial compliance. It was all or nothing, and a number of vendors chose “nothing.”

      and "The most recent version of the specification, OpenCL 2.2, was released back in 2017. Critically, it introduced the OpenCL C++ kernel language, finally bringing support for a more modern, object-oriented language to an API that was originally based on C. Equally critical however, three years later no one has adopted OpenCL 2.2. Not NVIDIA, not AMD, not Intel, and certainly not any embedded device manufacturer."

      https://www.anandtech.com/show/15746...ute-frameworks

      Comment


      • #33
        Originally posted by kieffer View Post
        Why can't we use OpenCL on ARM or PowerPC systems ? why is the driver version so restricted on those platforms ?
        I guess the same reason we can use OpenCL and CUDA on Windows and Linux, but not on FreeBSD - low userbase and relatively to that high support costs. IOW - just business.

        Comment


        • #34
          Originally posted by schmidtbag View Post
          Trust me, there are people here who will vehemently disagree with you on this, and suggest ECC is a necessity.
          That's not to say I feel ECC isn't necessary for Quadros, but having it as an option would be nice. In either case, I don't find this to be all that dubious - Nvidia isn't about to sell an expensive product and risk bit flipping compromising their reputation.
          When the calculation takes one millisecond and has been redundantly acquired, ECC is probably not needed (my case). It is not like when one re-uses those values in optimization routines ... Sometimes ECC is mandatory, sometimes it is not but for Nvidia, it is mandatory all the time to enforce you to purchase a 10k$ card when a 1k$ would as well do the job.

          Comment


          • #35
            Originally posted by blacknova View Post
            I guess the same reason we can use OpenCL and CUDA on Windows and Linux, but not on FreeBSD - low userbase and relatively to that high support costs. IOW - just business.
            That would explain the lack of support for PPC, but remember, Nvidia has the Tegra series. They also bought ARM. Surely, that would warrant their support.

            Originally posted by kieffer View Post
            When the calculation takes one millisecond and has been redundantly acquired, ECC is probably not needed (my case). It is not like when one re-uses those values in optimization routines ... Sometimes ECC is mandatory, sometimes it is not but for Nvidia, it is mandatory all the time to enforce you to purchase a 10k$ card when a 1k$ would as well do the job.
            Haha don't tell me, tell everyone else. I agree with you.

            Comment


            • #36
              Originally posted by kieffer View Post

              When the calculation takes one millisecond and has been redundantly acquired, ECC is probably not needed (my case). It is not like when one re-uses those values in optimization routines ... Sometimes ECC is mandatory, sometimes it is not but for Nvidia, it is mandatory all the time to enforce you to purchase a 10k$ card when a 1k$ would as well do the job.
              If you are talking about Quadro cards, then the main demographic for Quadro cards needs ECC since they rely on the computations that the GPU is being done is correct (think modelling for genetics, particles, physics etc etc).

              Also when talking about modern hardware, if implemented well ECC's performance impact is negligible. The only reason ECC itself is expensive because of artificial market segmentation, the reason why Quadro cards are so expensive is not mainly because of ECC but because NVidia knows that typical Quadro market can afford such a price.

              Also ECC is extremely handy if you are overclocking your GPU, its a very easy way to validate if you pushed the card too far. This is a good thing for gamers.
              Last edited by mdedetrich; 11 March 2021, 12:16 PM.

              Comment


              • #37
                OpenCL3.0 i think is actually necessery evil, because right now imagine you want to write program OpenCL.
                Your choices :
                -use 2.x fuck all hardware (that is a ton of it) that doesn't support 2.x,
                -use 1.2 and live in ancient world

                3.0 meanwhile gives you ability to use all 2.x features in more modern style, while at the same time if one particular hardware doesn't support one or 2 features you use, you can write small code path that handles it diffrently. and viola all works.

                OpenCL2.0 features:
                • Shared Virtual Memory
                • Dynamic Parallelism
                • Generic Address Space
                • Images
                • C11 Atomics
                • Pipes
                Now chance you want 1 or 2 features from it is high, but for example you don't need SVM. And 3.0 should allow you on PCs to support all of that except SVM, from all major hardware manufactures. Only problem is that:
                -OpenCL3.0 for a while will be gimstick because devs generally has to support older hardware then hardware manufactures update drivers,
                -it depends on hardware manufactures to be agressive in allowing optional features, when generally Nvidia, Intel and AMD are fine there, I have some worries in mobile GPU market.

                In nutshell I like OpenCl3.0 but... way too late. around 2-3 years too late.
                Last edited by piotrj3; 14 March 2021, 09:48 AM.

                Comment

                Working...
                X