Announcement

Collapse
No announcement yet.

Tiny Corp Changes Course Yet Again With Plans To Offer AMD Radeon GPUs

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

  • #51
    Originally posted by varikonniemi View Post
    and after that when people are not as easily fooled, they would reduced the price premium to what normally 24gb extra vram could sanely cost.
    the point is a "7900XTX 48gb" as you want it does not have ECC ram...

    means amd could produce a 7900XTX 48gb with reduced price as you want but then no ECC for you.

    AMD PRO W7900 means it has ECC ram.

    just keep in mind for the feature ECC alone you need 12.5% more chip die area ... what increase the cost but you want it for free and cheap of course.

    means it is impossible to sell a AMD PRO W7900 for the price of a hypothetical "7900XTX 48gb"...


    Phantom circuit Sequence Reducer Dyslexia

    Comment


    • #52
      Originally posted by novideo View Post
      Why can't AMD open up their firmware, and put all the proprietary features like DRM, HDMI 2.1, etc in an optional proprietary firmware blob/driver, like how Firefox makes DRM optional on Linux? And those that want DRM can just install the blob without making the firmware closed for everyone.
      they can't because of the requirements of the people who demand DRM/COPYPROTECTION and the requirements are that not only the High-bandwidth Digital Content Protection (HDCP)​ copyprotection module is a black box requirements say clear that the complete chip who handle High-bandwidth Digital Content Protection (HDCP)​ data is a black box.

      if AMD makes one open-source firmware and one who only has the closed source bits for the DHCP copyprotection then you only need to reverse engineer the small black box of the High-bandwidth Digital Content Protection (HDCP)​

      by make sure not only the High-bandwidth Digital Content Protection (HDCP)​ part is a black box instead the complete chip function is black box you increase the complexity to reverse engineer it.

      in the past companies like AMD could not ignore this because they did make more money with boxes play hollywood movie via bluray or netflix

      but this time has changed today they really make more money with AI and hollywood/bluray/netflix becomes less important.

      as bridgman told us in the past to release hardware with a open source firmware they would not only need to open-source it they by law would be forced to invent a complete new GPU chip architecture to make sure this information of the firmware source code does not help in any way to reverse engineer the HDCP copy protection.

      by law it is impossible to open the existing hardware or any hardware who supports HDCP copy protection.

      AMD could maybe open-source the firmware of the Instinct product line because HDCP is irrelevant for these compute cards.
      Phantom circuit Sequence Reducer Dyslexia

      Comment


      • #53
        Originally posted by Forge View Post
        At home I run AMD, because I like Linux and want GPUs as open as possible. At work we run Nvidia products, because they are in weight classes that AMD hasn't yet begun to explore. Show me an AMD product that can approach this: https://www.nvidia.com/en-us/data-ce...per-superchip/
        I hate it. My personal life has been so much better since I got Nvidia out of my Linux machines, but they are the only game in town at the highest end.
        isn't that not just a AMD instinct MI300A ??? https://www.amd.com/en/products/acce...00/mi300a.html
        Phantom circuit Sequence Reducer Dyslexia

        Comment


        • #54
          Originally posted by qarium View Post

          they can't because of the requirements of the people who demand DRM/COPYPROTECTION and the requirements are that not only the High-bandwidth Digital Content Protection (HDCP)​ copyprotection module is a black box requirements say clear that the complete chip who handle High-bandwidth Digital Content Protection (HDCP)​ data is a black box.

          if AMD makes one open-source firmware and one who only has the closed source bits for the DHCP copyprotection then you only need to reverse engineer the small black box of the High-bandwidth Digital Content Protection (HDCP)​

          by make sure not only the High-bandwidth Digital Content Protection (HDCP)​ part is a black box instead the complete chip function is black box you increase the complexity to reverse engineer it.

          in the past companies like AMD could not ignore this because they did make more money with boxes play hollywood movie via bluray or netflix

          but this time has changed today they really make more money with AI and hollywood/bluray/netflix becomes less important.

          as bridgman told us in the past to release hardware with a open source firmware they would not only need to open-source it they by law would be forced to invent a complete new GPU chip architecture to make sure this information of the firmware source code does not help in any way to reverse engineer the HDCP copy protection.

          by law it is impossible to open the existing hardware or any hardware who supports HDCP copy protection.

          AMD could maybe open-source the firmware of the Instinct product line because HDCP is irrelevant for these compute cards.
          Thanks, would you mind sending a link to bridgman's full post? Or giving a hint as to what article it was under?

          Comment


          • #55
            Originally posted by ET3D View Post
            It's like I'd narrate my software development, every attempt that doesn't work, and keep giving percentages to whether this will work in the end, while I keep being pessimistic about it.
            No, because what software you write is entirely within your control. The situation of AMD's GPU driver stack and tools are mostly outside Tiny Corp control.​

            Originally posted by Quackdoc View Post
            What part of the rocm stack that they are working with is closed source outside of the firmware? Ill keep asking, but so far no one has been able to tell me​​
            The problem is not the ROCm code, it is what is missing from ROCm and exists only as proprietary software.
            Originally posted by novideo View Post
            Only for the more obscure hardware that you mentioned, usually because the vendor wants it, just like PlayStation's and Xbox's proprietary AMD GPU drivers. For a consumer using an AMD GPU on their Linux PC using Mesa and/or ROCm, everything is open besides the firmware.
            You call it obscure but secondhand enterprise hardware was for a long time the only feasible and affordable way to run proper GPU compute on AMD hardware.

            AMD in their limitless disdain for home labs of course tried to make it as cumbersome as possible to get your hands on drivers if you don't have a business relationship with them.

            Comment


            • #56
              Originally posted by chithanh View Post
              The problem is not the ROCm code, it is what is missing from ROCm and exists only as proprietary software.
              Yes, and as far as I know, this is just the firmware of which Tinycorp is asking for the source for. If im wrong I wish someone would just tell me what it is already...

              Comment


              • #57
                Originally posted by qarium View Post

                isn't that not just a AMD instinct MI300A ??? https://www.amd.com/en/products/acce...00/mi300a.html
                I've had hands on with both. The GH100 beats the MI300 handily, and the GB200 is another leap ahead. AMD has good GPUs, that are decent GPGPUs, but nothing like the Grace chip, and no big, high-speed shared memory pool. Your MI300A rig has 128GB of high speed memory available to the chip. The GH100 has over 700GB available to theirs, and at comparable or better speeds. It's not fair, but that's why Nvidia is dominating the AI/GPU-HPC space for now.

                Comment


                • #58
                  Originally posted by novideo View Post
                  Thanks, would you mind sending a link to bridgman's full post? Or giving a hint as to what article it was under?
                  this is info from years many years ago. i would say its to much to give you the exact quote or even only the article. you can try google it but... its maybe better to just ask bridgman.

                  or else just trust my memory. you could try to google the requirements for HDCP but its very simple anything what HDCP touches in meaning of data need to be in black box outside of the black box anything need to be encrypted to be secure from COPY the data.

                  its plain and simple impossible to build a open hardware chip with open-source firmware with that requirement for HDCP

                  AMD can go in 2 ways for one they could say for AMD Instinct HDCP is not needed and make open-source firmware but keep in mind bridgman also told us as soon as you make opensource firmware the hardware is more or less also open hardware and AMD does not want open hardware because then china can copy it easily.

                  keep in mind the OpenPOWER solidsilicon.com people did exactly this because the OpenPOWER ISA of the power10 chip is opensource and the firmware is also open-source solidsilicon.com could just copy the complete chip.

                  if AMD makes the Instinct firmware open-source a china company can just copy the complete chip.

                  this means even if HDCP is out of the picture AMD still does not want their chips copies by chinese copy-cat companies.

                  Phantom circuit Sequence Reducer Dyslexia

                  Comment


                  • #59
                    Originally posted by Forge View Post
                    I've had hands on with both. The GH100 beats the MI300 handily, and the GB200 is another leap ahead. AMD has good GPUs, that are decent GPGPUs, but nothing like the Grace chip, and no big, high-speed shared memory pool. Your MI300A rig has 128GB of high speed memory available to the chip. The GH100 has over 700GB available to theirs, and at comparable or better speeds. It's not fair, but that's why Nvidia is dominating the AI/GPU-HPC space for now.
                    GH100 has only 80GB HBM2e VRAM
                    GB200 has only 141 GB HBM3E

                    the MI300A is 192GB of HBM3
                    Phantom circuit Sequence Reducer Dyslexia

                    Comment


                    • #60
                      Originally posted by qarium View Post

                      GH100 has only 80GB HBM2e VRAM
                      GB200 has only 141 GB HBM3E

                      the MI300A is 192GB of HBM3
                      That's on the GPU alone. Each GPU has 96GB of HBM. The CPU has 512GB of LPDDR5X. Any GPU can access any GPU's memory and the CPU's memory at 900GB/s. 512GB+96GB+96GB = the 700GB I was talking about.

                      The NVIDIA Grace Hopper Superchip Architecture is the first true heterogeneous accelerated platform for high-performance computing (HPC) and AI workloads. It accelerates applications with the…

                      Comment

                      Working...
                      X