Announcement

Collapse
No announcement yet.

Running Gallium3D's LLVMpipe On The Eight-Core 5GHz CPU

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

  • Running Gallium3D's LLVMpipe On The Eight-Core 5GHz CPU

    Phoronix: Running Gallium3D's LLVMpipe On The Eight-Core 5GHz CPU

    Having an eight-core CPU that can clock up to 5.0GHz (albeit having a 220 Watt TDP), curiosity got the best of me to run some quick (or slow) Gallium3D LLVMpipe tests just to see how this software fall-back driver performs...

    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
    considering the short pipelines and high frequency, I honestly expected better results. I'm a little disappointed, though I wonder if LLVMpipe is optimized for the bulldozer architecture and if so, can it take advantage of the modules? Because otherwise, this is basically only effectively running on 4 cores rather than 8.

    Comment


    • #3
      "At 1920 x 1080 for this old id Tech 3 game, the average frame-rate is only about 15 FPS. "
      This is incorrect, OpenArena is much more detailed than Quake 3 (as baseline for "old id Tech 3 game"), it can severely load gpu. For example p_vomit map is at least 20x as demanding as q3 maps.

      Comment


      • #4
        LLVMpipe finally can compete with Rage Fury MAXX

        Comment


        • #5
          lol

          better performance than fglrx

          Comment


          • #6
            I'm curious how fast it will be on 72-core Xeon Phi Knights Landing.

            Comment


            • #7
              Originally posted by schmidtbag View Post
              considering the short pipelines and high frequency, I honestly expected better results. I'm a little disappointed, though I wonder if LLVMpipe is optimized for the bulldozer architecture and if so, can it take advantage of the modules? Because otherwise, this is basically only effectively running on 4 cores rather than 8.
              llvmpipe can use up to 16 threads (actually you could change it to use more pretty easily though it might not scale all that well past 8, in particular because everything pre-rasterization is single threaded.
              But effectively for llvmpipe, a 4 module BD is really just a 4 core cpu - almost all of the (generated) code is to be executed by the simd unit. The very high latencies between the simd unit and the integer ones probably are not helping neither (because the little code that actually does run on the int units tends to be very dependent on the simd one), though I suppose in this case it helps that there 8 integer cores (as the simd unit can execute code from another thread meanwhile). There isn't all that much optimization I'm aware of that could be done specifically for BD (note that for the generated code llvm already does optimize for the specific cpu type).

              Comment


              • #8
                And I thought it would be about the Core i7-5960X. But that's "only" 3.0GHz-3.3GHz.

                A comparison with this CPU should be interesting, how much better power efficiency is.

                Comment


                • #9
                  Originally posted by Szzz View Post
                  I'm curious how fast it will be on 72-core Xeon Phi Knights Landing.
                  I doubt it would run very well. The single-thread bottleneck for geometry processing would likely get very serious. Plus there's zero optimization done for the vector instruction set of this cpu - theoretically llvm probably could generate some code for it but I highly doubt it would be very effective.
                  That said, AVX-512 would rock for emulating a gpu - proper built-in lane masking (just like gpus...), some instructions specifically tailored for meeting precision requirement of d3d10 (ok these are optional...), fast gather/scatter (the former being very helpful for texturing) and so on. But right now llvm doesn't even really use avx2.

                  Comment


                  • #10
                    In general we encourage the use of GPUs for floating-point-intensive workloads like... um... graphics
                    Test signature

                    Comment

                    Working...
                    X