Announcement

Collapse
No announcement yet.

Intel Mesa OpenGL Driver Lands 48-bit Addressing Support, Lets Up To ~256TB Of vRAM

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

  • Intel Mesa OpenGL Driver Lands 48-bit Addressing Support, Lets Up To ~256TB Of vRAM

    Phoronix: Intel Mesa OpenGL Driver Lands 48-bit Addressing Support, Lets Up To ~256TB Of vRAM

    Intel's i965 Mesa OpenGL driver now allows for 48-bit addressing, which greatly expands the GPU memory limits...

    Phoronix, Linux Hardware Reviews, Linux hardware benchmarks, Linux server benchmarks, Linux benchmarking, Desktop Linux, Linux performance, Open Source graphics, Linux How To, Ubuntu benchmarks, Ubuntu hardware, Phoronix Test Suite

  • #2
    I can't wait to see this feature implemented in the amdgpu driver too...

    Comment


    • #3
      So, usecases? iGPU in a server with massive amounts of RAM that can be allocated as vRAM?

      Comment


      • #4
        My guess is for getting OpenCL working with huge datasets, or maybe virtual servers where each one gets a big slice of vram backed by this

        Comment


        • #5
          Let's say I want to buy an Intel video card with 256TB of vRAM. How do I go about it?

          Comment


          • #6
            Originally posted by eydee View Post
            Let's say I want to buy an Intel video card with 256TB of vRAM. How do I go about it?
            Work for the NSA?

            Comment


            • #7
              Such cynicism. Look: just like CPU DRAM 20 years ago, now you gotta expand video memory addressing beyond 32bit's 4GB. What's the next reasonable step? FWIW, Nvidia CUDA 8 on Pascal+ virtual address space is 49 bits -- but don't ask how they got there.

              Comment


              • #8
                Originally posted by eydee View Post
                Let's say I want to buy an Intel video card with 256TB of vRAM. How do I go about it?
                You can use regular disks as vRAM I think. There is a professional AMD GPU that takes NVMe SSDs for up to 2TB vRAM iirc. It's considerably slower than using actual video memory, but has it's uses when you need the large amount of memory. At my last job it would have been handy as using 12GB of vRAM on Titans was easy to max, as is the 128GB RAM we had, but EPYCs were out of stock at the time for affordable machines that could exceed 128GB RAM(We went with ThreadRipper and easily maxed the 16 cores, 32 threads, unfortunately it doesn't support LR-DIMMs, have to wait for UDIMMs to increase in density).

                256TB RAM is the thing of super computers :P Large clusters, hella expensive.

                Comment


                • #9
                  Nice thing that GPUs with 128TB+ vRAM are dirt cheap now /s

                  Originally posted by polarathene View Post
                  You can use regular disks as vRAM I think. There is a professional AMD GPU that takes NVMe SSDs for up to 2TB vRAM iirc. It's considerably slower than using actual video memory, but has it's uses when you need the large amount of memory.
                  As far as I know, another advantage is that you have dedicated bandwidth to the SSD, so you can still use the PCI lanes for normal operations instead of just swapping in and out the needed memory.
                  Last edited by M@yeulC; 02 March 2018, 11:07 AM.

                  Comment


                  • #10
                    This reminds me of https://xkcd.com/619/
                    But really it's probably not that far fetched that this feature has some use and maybe not today but possibly tomorrow (don't think it too concretely).

                    BTW, funny that the title text talks about Intel cards. We haven't had those in two decades I think.
                    Last edited by Tomin; 02 March 2018, 11:44 AM.

                    Comment

                    Working...
                    X