Sujet Les questions auxquelles on n'a pas les réponses mais un AFien peut les avoir

Citation :

I just released 3.7.16.  There's some misc stuff, but the big news is

 that I have added architectural support for the Core2 family (though have

 not tuned any assembly for it yet), and this guy is a beast!  Looks like

 after years of getting schooled by AMD, Intel finally got tired of it.  Here's

 a comparison of flops/cycle comparing Core2 vs Athlon-64:

              

              SCALR x87  SCALR SSE  VECTR SSE

              ---------  ---------  ---------

              sngl dble  sngl dble  sngl dble

 Core2           1    1     2    2     8    4

 Athlon-64       2    2     2    2     4    2

 

 This means Core2 is the fastest floating point x86 ever produced.  It ties

 the G5, which was previously the best peak performer of any desktop system.

 I think they added an extra vector unit to Core2 vs Core1 to get this

 doubling.  Note that even using Scalar SSE, you can get twice the clock

 rate on the Core 2 (unlike prior Intel machines).  So, from this we know

 that computationally, the only place AMD is winning is x87 code.  This

 actually gives the AMD a double advantage for scalar code (right now

 all C code results in scalar performance): (1) since the x87 has same

 peak as SSE, you can use it for greater accuracy;  (2) x87 actually

 has a slight performance advantage in practice, since in the CISC

 instruction set, x87 code takes up a lot less instruction cache.

 In older archs, AMD unambiguously won for scalar code, but with the extra

 vector unit, this is no longer so unambiguous.  On the other hand, Core2

 should dominate anytime memory isn't the only cost (I haven't really timed

 memory stuff, but it doesn't look like AMD is dominating there either).  

 

 Those are peaks, so what are we getting in practice?  For reference here's

 my 2.2Ghz Athlon-64:

 

 *******************************************************************************

 Clock rate=2200Mhz

                single precision        double precision

             *********************    ********************

                real      complex       real      complex

 Benchmark   %   Clock   %   Clock   %   Clock   %   Clock

 =========   =========   =========   =========   =========

   kSelMM       354.1      339.8      161.7      180.0

   kGenMM       185.2      156.2      161.7      169.2

   kMM_NT       136.8      146.6      113.6      132.3

   kMM_TN       151.6      142.0      132.2      147.2

   BIG_MM       338.3      330.3      159.6      174.5

    kMV_N        53.8      138.8       35.9       72.7

    kMV_T        61.3       73.0       33.5       52.5

     kGER        44.9       91.8       23.7       48.0

 *******************************************************************************

 

 Now here's my 2.4Ghz Core2Duo:

 *******************************************************************************

 Clock rate=2394Mhz

                single precision        double precision

             *********************    ********************

                real      complex       real      complex

 Benchmark   %   Clock   %   Clock   %   Clock   %   Clock

 =========   =========   =========   =========   =========

   kSelMM       531.8      517.8      270.1      285.8

   kGenMM       174.7      168.0      165.3      165.3

   kMM_NT       141.8      141.0      133.5      136.3

   kMM_TN       158.8      159.3      142.9      150.1

   BIG_MM       501.0      501.0      262.5      288.4

    kMV_N       110.6      202.5       57.8       95.5

    kMV_T        91.3      101.5       49.9       84.7

     kGER       153.9      113.5       28.9       46.4

 *******************************************************************************

 

 So, there is no question that the Core2 gives the Athlon-64 a drubbing pretty

 much across the board, as we would expect.  Note that I have not yet

 optimized any of the kernels for the Core2: this is just using the

 preexisting kernels (mostly tuned for netburst)!  I believe we have a decent

 amount of room on the floor here.  In particular, the present kernels do not

 exploit the doubling of registers.  I think the computational ordering is

 probably OK, but the prefetch probably should be tuned, which might also

 drift the NB . . .  So, I think these numbers will eventually become even

 more dominant.

 

 So, for now, Core2 is the absolute, unambiguous king of x86 floating point.

 I very much hope AMD can respond!

Citation : Prends un mac

Citation : Ce fichier XML ne semble pas avoir d'information de style lui étant associé. L'arbre du document est montré ci-dessous.