2 Replies Latest reply on Jun 6, 2013 1:19 PM by kd2

    Theoretical peak FLOPS

    niederb

      I have question regarding the theoretical peak FLOPS of my graphics card.

      I have a Radeon HD 7970 GHz Edition which has a peak of 4096 GFLOPS.

       

      Now I would wanted to calculate this number.

      When I just multiply the core clock (1000 MHz) with the number of cores (2048) I get 2048 GFLOPS.

       

      What am I missing to get the 4096 FLOPS? Does a MAD operation count as two operations?

      Then again according to the AMD Accelerated Parallel Processing OpenCL Programming Guide it can execute 4 MAD, ADD or MUL operations per cycle (Page 5-23)?

      Which would mean the number of FLOPS would be even higher.

       

      Can someone explain how to calculate the theoretical peak FLOPS?

       

      Many thanks!

      Niederb

        • Re: Theoretical peak FLOPS
          nou

          yes MAD is counted as two operations. one cycle is executed in four clock as it executes 64 wide wavefront in four clocks.

          • Re: Theoretical peak FLOPS
            kd2

            Part of the ever-existing confusion is that the terminology is dependent on context. When talking about the Southern Islands / GCNs, the 7970 has 32 Compute Units (CUs), each CU has 4 SIMDs, and each SIMD having 16 Processing Elements (ALUs).  So simple.. with a FMA counting as 2 ops/cycle, that gives you 4096 flops. But some of the documentation in the Southern Islands docs (like the table on 5-23) is more structured as if it was still talking about the Evergreen and Northern Islands chips where the Compute Unit is described as 16 Stream Processors (SP) and the SP does 4 ops per cycle.