Announcement

Collapse
No announcement yet.

Some benches on the upcoming 45nm AMD

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

  • #21
    Originally posted by MetalheadGautham View Post
    Honestly, I think they suck. NO AM3, no HT 3.0 and NO DDR3.
    Eh? HT3.0 is already part of AM2+.

    And the AM2+ processors all have DDR3-capable memory controllers, it's just that nobody has made any DDR3 motherboards for them. And as far as I can tell, there's really not much point at the moment.

    And Core i7 is waaaay ahead of it. Multimedia performance is so low compared to Yorkfield and Bloomfield. And AMD has nothing to combat Intel's HT Technology.
    i7 is definitely far ahead. Nothing else to say about that.

    As for multimedia, I'm still suspicious of the compiler technology; architecturally K10's SSE support looks really good on paper. It seems to me that we just haven't seen code that was generated properly to take full advantage of it. And HT really doesn't buy anything if you can keep your active cores sufficiently fed that they have fewer stalls.

    Comment


    • #22
      Originally posted by highlandsun View Post
      Eh? HT3.0 is already part of AM2+.

      And the AM2+ processors all have DDR3-capable memory controllers, it's just that nobody has made any DDR3 motherboards for them. And as far as I can tell, there's really not much point at the moment.
      Actually the AM2+ Phenom II's have the DDR3 mem controllers disabled. You will never see a AM2+ Phenom II with DDR3. I actually would think that the AM2+ Phenom II's are actually early revisions with faulty DDR3 controllers but their DDR 2 controllers worked fine. The AM2+ Phenom II's will never work in a AM3 board because of this (same with any other AM2+ processor). Your right though HT 3 has been around for a long time though already. Your also correct that DDR3 at this point doesn't offer much, if any, in real world performance.


      i7 is definitely far ahead. Nothing else to say about that.

      As for multimedia, I'm still suspicious of the compiler technology; architecturally K10's SSE support looks really good on paper. It seems to me that we just haven't seen code that was generated properly to take full advantage of it. And HT really doesn't buy anything if you can keep your active cores sufficiently fed that they have fewer stalls.
      Yes, the i7 is one heck of a chip, it also carries one hell of a premium price tag as well with needing to have a new MB, expensive ram, etc. HT as you said, only can show any gain if the core's are not under full load. Under full load performance may actually slightly decrease.

      Now onto SSE. Intels SSE does have a advantage in it's hardware. The intels cores have 3 128-bit units with 2 of them being symmetric vs the old K8's 2 64-bit units. With Phenom SSE performance did see an increase though with them now finally being able to execute two 128-bit generic SSE ops and 1 SSE MOV per cycle. Still the intels do carry a bit of an advantage here with it's 3 128-bit units.

      Compilers do make a difference as well but greater gains can actually be had by optimizing the code for a specific processor that compilers may try to do but don't always succeed at. GMP for example when optimized with AMD specific patches abolutely will trounce on a intel with simular hand optimization (to be fair though how the i7 would respond isn't really known yet). Matlab, SuperPi, Prime95 is what intel guys love to show off for intel's "mathematical superiority" but all of those carry very intel specific optimizations or genaric optimizations without the same level of optimization done for the AMD specific CPU's.

      Comment


      • #23
        I did some first PTS tests with the Phenom II X4 940 today.



        Nexuiz, OpenSSL and 7zip. Nexuiz in different settings (drivers, kernels).

        What I found out so far: performance is awesome, CnQ does not slow down, 790GX/SB750 work fine with Linux although I used a HD 4850 1GB.

        System was Debian/sid.

        Comment


        • #24
          nexuiz can not use 4 cores, so a intel dual core with 2.4 ghz beats it (amd quad running at 3.0 ghz!), at 3.2 ghz even 33% - that bench scales really good from 9*266 to 8*400. Or maybe fglrx is just too slow

          Comment


          • #25
            Some basic bandwidth and FP number crunching tests I ran on Opteron 2384 and Xeon 5450



            In aggregate, 8 cores fully loaded, the Opteron still wins, but the Xeons win most of the single-thread runs. There's something really weird at the 2Kx2K size where the Xeon performance drops. Perhaps a bug in the FFTW library.

            Also an interesting note - the code compiled with gcc -march=core2 runs slower on Opteron than code compiled for -march=amdfam10. Kind of expected. But the code compiled for -march=amdfam10 ran *faster* on the Xeon than the code compiled for -march=core2. gcc 4.3 is doing something weird...

            Comment


            • #26
              Originally posted by highlandsun View Post
              Some basic bandwidth and FP number crunching tests I ran on Opteron 2384 and Xeon 5450



              In aggregate, 8 cores fully loaded, the Opteron still wins, but the Xeons win most of the single-thread runs. There's something really weird at the 2Kx2K size where the Xeon performance drops. Perhaps a bug in the FFTW library.

              Also an interesting note - the code compiled with gcc -march=core2 runs slower on Opteron than code compiled for -march=amdfam10. Kind of expected. But the code compiled for -march=amdfam10 ran *faster* on the Xeon than the code compiled for -march=core2. gcc 4.3 is doing something weird...

              GCC's intel optimization has never been all that great. GCC on amd's on the other hand enjoy great architecture optimizations in GCC / nasm / yasm/ etc. Not surprising as intel wishes for people to purchase their compilers instead.

              Comment

              Working...
              X