Announcement

Collapse
No announcement yet.

OpenCL 3.0.16 Released With One New Extension, Semaphores & External Memory Finalized

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

  • #21
    Originally posted by qarium View Post
    there are already people who post benchmarks between openCL and Vulkan Compute and in their micro bechmarks vulkan won big...
    It's not intrinsic to Vulkan. It has to do with the relative quality of the drivers and how well-optimized the code is, in each case.

    Originally posted by qarium View Post
    ​its even worst relevant apps like Blender did even drop OpenCL completely because it never worked for complex projects.
    Blender is known to be developed by Nvidia fanboys. They tolerate other backends, but don't maintain them. The only reason there's a HIP/ROCm backend and a oneAPI backend is because AMD and Intel both wrote and maintain them. OpenCL has no similar backer, so of course it's going to fall into disrepair, and once that happens people will stop even trying to use it and the situation will just spiral.

    Originally posted by qarium View Post
    ​​also there are C vs rust benchmarks who proof that rust outperforms C and well OpenCL more or less use a C dialect
    Again, you have no idea what you're talking about.

    Originally posted by qarium View Post
    ​​​about FPGA you know that there is a professor who released a AMD ISA open-source implementation for FPGAs to emulate a AMD GPU??
    That's never going to be as efficient as compiling OpenCL directly for the FPGA.

    Comment


    • #22
      Originally posted by coder View Post
      It's not intrinsic to Vulkan. It has to do with the relative quality of the drivers and how well-optimized the code is, in each case.
      Blender is known to be developed by Nvidia fanboys. They tolerate other backends, but don't maintain them. The only reason there's a HIP/ROCm backend and a oneAPI backend is because AMD and Intel both wrote and maintain them. OpenCL has no similar backer, so of course it's going to fall into disrepair, and once that happens people will stop even trying to use it and the situation will just spiral.

      but the point is the Vulkan-Compute and Vulkan-Graphic+hardware raytracing backends for Blender is not dropped and is actively under development.
      thats the point OpenCL has no backer at all.. but people back Vulkan-Compute and Vulkan-Raster/raytracing...

      also keep in mind many ARM SOCs and graphic cards outside of Nvidia/AMD/Intel will never get ROCm/HIP or CUDA or OneAPI support
      This means they are all on Vulkan-Compute and Vulkan-Raster/raytracing... thats the reason why they develop this for Blender.

      you really look sad that OpenCL vanish but in years if Vukan-Compute and Vulkan-Raytracin emerge in Blender and other apps no one will miss OpenCL...

      Originally posted by coder View Post
      Again, you have no idea what you're talking about.
      That's never going to be as efficient as compiling OpenCL directly for the FPGA.
      remember first you said it does not run on FPGA then i showed to you it runs on FPGA now you say its not efficient.

      and no one stops anyone to write a Vulkan-compute to FPGA compiler to autogenerate efficient code for such FPGA chips
      Phantom circuit Sequence Reducer Dyslexia

      Comment


      • #23
        Originally posted by qarium View Post
        but the point is the Vulkan-Compute and Vulkan-Graphic+hardware raytracing backends for Blender is not dropped and is actively under development.
        The main area where Vulkan is used by Blender is for interactive rendering, which is fundamentally different than the photorealistic renderer used for final output.

        Originally posted by qarium View Post
        OpenCL has no backer at all..
        Perhaps in the narrow context of Blender, but you can't just extrapolate from that one example.

        Originally posted by qarium View Post
        keep in mind many ARM SOCs and graphic cards outside of Nvidia/AMD/Intel will never get ROCm/HIP or CUDA or OneAPI support
        You're wrong about that last point. Google, Intel, and Qualcomm just announced an alliance around pushing oneAPI as an alternative to CUDA. oneAPI has OpenCL at its foundation.

        Originally posted by qarium View Post
        remember first you said it does not run on FPGA then i showed to you it runs on FPGA now you say its not efficient.
        To be viable as a compute accelerator, FPGAs need to be efficient! If you show me an option that doesn't efficiently use the FPGA's resources, I say that's no option at all!

        Originally posted by qarium View Post
        ​no one stops anyone to write a Vulkan-compute to FPGA compiler to autogenerate efficient code for such FPGA chips
        As I said, Vulkan doesn't provide a language for the code actually running on the device. What Vulkan does is provide the host API, and it's no use implementing Vulkan support for a device which cannot even meet all of Vulkan's core requirements.

        Again, you obviously don't know what you're talking about. Please stop wasting my time & yours.

        Comment


        • #24
          Originally posted by coder View Post
          The main area where Vulkan is used by Blender is for interactive rendering, which is fundamentally different than the photorealistic renderer used for final output.
          if you do a Blender Vulkan renderer for interactive rendering who exactly stops you to turn it into a final photorealistic renderer with vulkan-compute and vulkan-raytracing ??...

          Originally posted by coder View Post
          Perhaps in the narrow context of Blender, but you can't just extrapolate from that one example.
          why? other projects did abolish openCL to...

          Originally posted by coder View Post
          You're wrong about that last point. Google, Intel, and Qualcomm just announced an alliance around pushing oneAPI as an alternative to CUDA. oneAPI has OpenCL at its foundation.
          i am really not used to to get fooled by you because now you double your argument oneAPI is from intel and you say "Google, Intel, and Qualcomm just announced an alliance around pushing oneAPI" so its only google and intel joining oneAPI...

          but keep in mind that all these players google, intel, qualcomm are the losers in the AI and Compute field they lose big against Nvidia and also AMD,..

          if you watch the company numbers on intel they are big big losers they lost 6,8 billion euro on ARC GPUs alone because their performance in the raster graphic and raytracing graphic field outside of the 155 notebook chip is a flop.

          right now intel also lose in the AI field because they invest big in it but yet have no cashflow coming in because only high-risc investors buy intel hardware instead of cuda nvidia hardware.

          many in the field better go with ROCm/HIP and zluda to emulate CUDA...

          keep in mind intel loses big money on process nodes to they build EUV based factories who costs 30 billion euro and yet not a single EUV factory is ready to produce their chips thats the reason why they produce chips at TSMC...

          keep in mind that Google lost big in the AI market and they literally lost the AI race no one can take them serious

          and qualcomm also is not known for AI solutions outside of some small NPU(6/8bit minifloat) units on their mobile ARM SOCs

          Originally posted by coder View Post
          To be viable as a compute accelerator, FPGAs need to be efficient! If you show me an option that doesn't efficiently use the FPGA's resources, I say that's no option at all!
          I really don't know why FPGAs have relevance for you because all the AI only accelerators are NPUs (neural processing unit​) with minifloat hardware or Matrix cores like the AMD:

          "Matrix-Core-Technologien, die erweiterten Rechendurchsatz mit verbesserter Parallelverarbeitung auf Anweisungsebene bieten. Hierzu gehört die Unterstützung für eine Vielzahl von Genauigkeitsstufen (INT8, FP8, BF16, FP16, TF32, FP32 und FP64) sowie Sparse-Matrix-Daten"

          right now FPGAs have zero relevance in the AI and Compute field...

          but if you see the AMD XILINX FPGA​ they have tools and compilers to compile your program code into a FPGA grid...

          Originally posted by coder View Post
          As I said, Vulkan doesn't provide a language for the code actually running on the device. What Vulkan does is provide the host API, and it's no use implementing Vulkan support for a device which cannot even meet all of Vulkan's core requirements.
          to be honest you talk bullshit here because like with CDNA Instinct accelerators who do not support Vulkan Core requierements they do support the Vulkan-Compute requirements ...

          also keep in mind that vulkan-compute is also much more than you need for AI with AI you could have a even smaller feature set.

          Originally posted by coder View Post
          Again, you obviously don't know what you're talking about. Please stop wasting my time & yours.
          no one is forcing you to read my posts and no one is forcing you to write answers to me.
          you still can not explain why literally the complete compute and ai market is on CUDA and not on oneAPI...

          I honestly would not even waste my time in learning OpenCL or oneAPI i would better learn vulkan and rust and do vulkan-rust-compute..
          Phantom circuit Sequence Reducer Dyslexia

          Comment


          • #25
            Not replying to the other stuff, because it's veering off-track.

            Originally posted by qarium View Post
            right now FPGAs have zero relevance in the AI and Compute field...
            Not true. FPGAs are very complementary to other forms of compute acceleration. They're very good at functional parallelism and where low-latency and power-efficiency are at a premium. Embedded signal processing remains a classic example.

            Yet again, you trip over your own ignorance.

            Originally posted by qarium View Post
            ​to be honest you talk bullshit here because like with CDNA Instinct accelerators who do not support Vulkan Core requierements
            Which ones?

            Originally posted by qarium View Post
            ​​you still can not explain why literally the complete compute and ai market is on CUDA and not on oneAPI...
            I'll give you 3:
            1. CUDA beat oneAPI to market by about 10 years
            2. oneAPI hardware (other than CPUs) comprises just a tiny share of the market
            3. there's no oneAPI hardware as fast as the CUDA GPUs out there
            That can change. But, if you're looking backwards, then it was obvious oneAPI was a long-term play and not something that could displace CUDA in the short term.

            Originally posted by qarium View Post
            ​I honestly would not even waste my time in learning OpenCL or oneAPI i would better learn vulkan and rust and do vulkan-rust-compute..
            This is a fraudulent statement, since you don't actually learn any APIs. You just post on here about them & other things you don't have any experience with.

            Comment


            • #26
              Originally posted by coder View Post
              Not replying to the other stuff, because it's veering off-track.
              Not true. FPGAs are very complementary to other forms of compute acceleration. They're very good at functional parallelism and where low-latency and power-efficiency are at a premium. Embedded signal processing remains a classic example.
              Yet again, you trip over your own ignorance.
              the mainstream does not add FPGA's right now all the cpu and GPUs and SOCs only add NPUs and ASIC and Matrix cores.
              no one adds FPGA's so the people who buy FPGA mostly do prototyping or build hardware like mega65 who will never go out of prototyping status.

              "They're very good at functional parallelism and where low-latency and power-efficiency are at a premium. Embedded signal processing remains a classic example."

              thats maybe true but its a niche and no mainstream hardware adds this. all the mainstream SOCs or GPUs or NPUs do not have FPGAs ...

              Originally posted by coder View Post
              Which ones?
              these Vega20 chips where the last one based on real 3D GPUs all the Instinct chips after Vega...

              they support Vulkan-compute but not vulkan-3D-Rasterizing also not Vulkan-hardware raytracing.

              Originally posted by coder View Post
              I'll give you 3:
              1. CUDA beat oneAPI to market by about 10 years
              2. oneAPI hardware (other than CPUs) comprises just a tiny share of the market
              3. there's no oneAPI hardware as fast as the CUDA GPUs out there
              That can change. But, if you're looking backwards, then it was obvious oneAPI was a long-term play and not something that could displace CUDA in the short term.
              you are right but this shows that in the market right now oneAPI has no better position than Vulkan-Compute
              its more or less a fact that both have very bad position.
              and that AMD is not on the oneAPI shit shows that something is fishy more or less AMD thinks its not a winning strategy

              Originally posted by coder View Post
              This is a fraudulent statement, since you don't actually learn any APIs. You just post on here about them & other things you don't have any experience with.
              how exactly can a personal opinion about what i want do in my free time be fraudulent ? it plain and simple can not be fraudulent.

              i tell you what i do and i do extactly what i tell you right now i read my CompTIA Security+ Certification Guide its already 70% done reading and i wait for my ordered german Rheinwerk Rust Handbook. i already have a english rust book but i as a german i more or less need to read 2 books one in german and one in english

              https://www.amazon.de/Rust-deutschsp...C5%BD%C3%95%C3 %91&crid=1U8HFF9F0KYBI&dib=eyJ2IjoiMSJ9.Pn2hyk7Naa wO7RlzxHLtYCBuvLWXE63-Pb5sBFDlBBaLN0YecGZdpzkeKcjr12j5Y58Yyn9KWxszy-InvN3A6ufd-xkdG3amAOTuzgs8tKLXDYywUbpWNQY0ecwiv5nhVrTPgXtuYfn-PUUIsJ0YWg.jhGS6LfLZEDbSoWxqMUVCmbPLpD6Itkk9jvK_xg KTzE&dib_tag=se&keywords=rheinwerk+rust&qid=171320 3793&sprefix=rheinwerk+rust%2Caps%2C125&sr=8-1

              i will also order books about Vulkan ... but i tell you what i will not do i will not buy books about CUDA and i will not buy books about oneAPI...

              there is clearly no possibility in any shape or form that my statement is a fraudulent statement ... a opinion or action of a person can not be a fraudulent statement...

              and to be honest i do not know why you perform hostility against me in your writing.

              for whatever reason you have invested time and resources in OpenCL and oneAPI and for you its a big loss if OpenCL does not become the future standard.
              Phantom circuit Sequence Reducer Dyslexia

              Comment


              • #27
                Originally posted by qarium View Post
                the mainstream does not add FPGA's right now
                ...
                its a niche and no mainstream hardware adds this. all the mainstream SOCs or GPUs or NPUs do not have FPGAs ...
                You should look up how much AMD paid for Xilinx and what their revenues were. Xilinx had much higher revenues that AMD ever got from GPUs.

                Again, you don't know what you're talking about... and you keep talking!

                Originally posted by qarium View Post
                ​this shows that in the market right now oneAPI has no better position than Vulkan-Compute
                Oh, it does, and part of that has to do with the fact that it's not Vulkan! You think technologies are just a shell game and you can just swap in whichever one you want... that's because you view them like football teams or brands, and not like tools that have real practical implications.

                Originally posted by qarium View Post
                ​​and that AMD is not on the oneAPI shit shows that something is fishy more or less AMD thinks its not a winning strategy
                AMD's "winning strategy" is just to copy CUDA. It's funny how you use AMD's judgment to support your arguments when it suits you, but disregard it when it doesn't.

                That's because you're nothing more than a fanboy. You don't know or understand the tech, so details seem irrelevant to you. You are just backing one vs. another for your own non-technical reasons.

                Originally posted by qarium View Post
                ​​​how exactly can a personal opinion about what i want do in my free time be fraudulent ? it plain and simple can not be fraudulent.
                You're giving advice as though you know what you're talking about. It'd be like someone who's a fan of legal drama television programs trying to give actual legal advice, without ever having spent even a day in law school.

                Originally posted by qarium View Post
                ​​​​i will also order books about Vulkan ...
                Good luck, then. Let us know when you release your first Vulkan project.

                Originally posted by qarium View Post
                ​​​​​and to be honest i do not know why you perform hostility against me in your writing.
                A person who persists in arguing a position without understanding the merits of the case is little more than a nuisance. You're talking as if you have some expertise in GPU compute programming, but you don't. In 1980's skateboard culture, they called people like you a "poser". You try to make up in quantity what your arguments lack in quality, but piling one bad point after another never adds up to a good argument.

                I don't expect you actually care about these points, but if you want to understand why people get frustrated with you and lash out, then maybe you should think about them.

                I'm not saying you shouldn't have opinions and preferences or that you shouldn't say them, but it would be wise to learn to recognize when you are starting to make claims about things you really don't know in real depth and just stop. It's not enough simply to have good intentions. If you don't really know what you're talking about, then trying to argue a point could actually do more harm than good.

                Comment


                • #28
                  Originally posted by coder View Post
                  You should look up how much AMD paid for Xilinx and what their revenues were. Xilinx had much higher revenues that AMD ever got from GPUs.
                  its right xilinx makes a ton of money but not with consumers in any shape or form. i am pretty sure most of their FPGA goes into companies who perform prototyping with that. means they develop hardware with it and simulate this on the FPGAs before it goes into production.

                  i still do not know any xilinx product outside the 6/8bit minifloat NPU they developed who landet in the amd cpus
                  but this NPU is not a FPGA ..

                  Originally posted by coder View Post
                  Again, you don't know what you're talking about... and you keep talking!
                  Oh, it does, and part of that has to do with the fact that it's not Vulkan! You think technologies are just a shell game and you can just swap in whichever one you want... that's because you view them like football teams or brands, and not like tools that have real practical implications.
                  AMD's "winning strategy" is just to copy CUDA. It's funny how you use AMD's judgment to support your arguments when it suits you, but disregard it when it doesn't.
                  That's because you're nothing more than a fanboy. You don't know or understand the tech, so details seem irrelevant to you. You are just backing one vs. another for your own non-technical reasons.
                  i am pretty sure that a copy of CUDA has higher marketshare than OneAPI...

                  Originally posted by coder View Post
                  You're giving advice as though you know what you're talking about. It'd be like someone who's a fan of legal drama television programs trying to give actual legal advice, without ever having spent even a day in law school.
                  there are people who can not give a advice that's right but you are the kind of person who never can get a advice.

                  Originally posted by coder View Post
                  Good luck, then. Let us know when you release your first Vulkan project.
                  thats a good start a even better one would be to team up to make it happen.

                  Originally posted by coder View Post
                  A person who persists in arguing a position without understanding the merits of the case is little more than a nuisance. You're talking as if you have some expertise in GPU compute programming, but you don't. In 1980's skateboard culture, they called people like you a "poser". You try to make up in quantity what your arguments lack in quality, but piling one bad point after another never adds up to a good argument.
                  I don't expect you actually care about these points, but if you want to understand why people get frustrated with you and lash out, then maybe you should think about them.
                  I'm not saying you shouldn't have opinions and preferences or that you shouldn't say them, but it would be wise to learn to recognize when you are starting to make claims about things you really don't know in real depth and just stop. It's not enough simply to have good intentions. If you don't really know what you're talking about, then trying to argue a point could actually do more harm than good.
                  i do my best to catch up a little.
                  Phantom circuit Sequence Reducer Dyslexia

                  Comment


                  • #29
                    Originally posted by qarium View Post
                    its right xilinx makes a ton of money but not with consumers in any shape or form. i am pretty sure most of their FPGA goes into companies who perform prototyping with that. means they develop hardware with it and simulate this on the FPGAs before it goes into production.
                    No, the way you make those kinds of revenues is by selling in high volumes. Prototyping is a very low-volume business. Again, you have no clue...

                    Originally posted by qarium View Post
                    ​i still do not know any xilinx product outside the 6/8bit minifloat NPU they developed who landet in the amd cpus
                    If you had any legitimate claim of expertise, then your familiarity with their products (or lack thereof) might be a useful data point.

                    However, as I've said time and again, you're out of your depth. Your ignorance of their products says nothing more than that you're unqualified to be trying to have this discussion.

                    Originally posted by qarium View Post
                    there are people who can not give a advice that's right but you are the kind of person who never can get a advice.
                    I don't listen to you because you're trying to tell me about something you obviously don't know nearly as well as I do. I do indeed listen to people whom I respect, but you've done nothing to earn that respect and a lot to show you deserve no benefit of the doubt.

                    Comment


                    • #30
                      Originally posted by coder View Post
                      No, the way you make those kinds of revenues is by selling in high volumes. Prototyping is a very low-volume business. Again, you have no clue...
                      yes right high volumes but can you show me any consumer products with xilinx FPGA inside ? because i have never seen one in my life.

                      Originally posted by coder View Post
                      If you had any legitimate claim of expertise, then your familiarity with their products (or lack thereof) might be a useful data point.
                      However, as I've said time and again, you're out of your depth. Your ignorance of their products says nothing more than that you're unqualified to be trying to have this discussion.
                      just show me the products because if i go in a hardware store for consumers i can not see any xilinx FPGA products.

                      Originally posted by coder View Post
                      I don't listen to you because you're trying to tell me about something you obviously don't know nearly as well as I do. I do indeed listen to people whom I respect, but you've done nothing to earn that respect and a lot to show you deserve no benefit of the doubt.
                      just tell me for strategic future planing do i even need to know the low-level stuff you know ? the simple answer is no.
                      in this meaning from a meta future planing strategic standpoint OpenCL/OneAPI will go exactly nowhere.
                      the stackeholders who support this route: Intel/Google/qualcomm are not in the position to make it happen.
                      they all Intel/Google/qualcomm lost the AI and compute war already and all they do is damage control to mitigate their loss.

                      in an arena of arguments the necessary of respect is a biased concept. to unterstand or to write a good argument no one even need respect.

                      "you've done nothing to earn that respect"

                      I honestly do not even want your respect ? whats the point of having your respect ?

                      "and a lot to show you deserve no benefit of the doubt."

                      who cares ? its your psychological problem not my one.
                      Phantom circuit Sequence Reducer Dyslexia

                      Comment

                      Working...
                      X