Try 474kn/s on 2.5GHz *XP*, 430kn/s on a A64 3500+
By mod on Wednesday, August 11, 2004 5:22 AM EDT
--------------------------------------------------------------------------------
Anands binaries for the A64 are next to completely unoptimized. His Athlon 64 3500+ somehow got 320kn/s, the one I tested got 430kn/s. His UBench and JTR scores are way lower as well. I don't have my Ubench results on hand but I can easily say that they were much higher than what Anand got on the 3500+ when using my Athlon XP.
For TSCP my compiler options were:
-O3 -march=athlon-xp -funroll-loops -fomit-frame-pointer -ffast-math -fprofile-arcs
Ran the benchmark, recompiled with:
O3 -march=athlon-xp -funroll-loops -fomit-frame-pointer -ffast-math -fbranch-probabilities
Binaries are available from my FTP:
ftp://ftp.newageoc.com/pub/bench/tscp-k7.zip (Windows - MingW32 gcc 3.4.1)
ftp://ftp.newageoc.com/pub/bench/tscp-k7lnx.zip (Linux - GCC 3.4.1)
Here are the results I've done with TSCP 1.81 so far.
Athlon XP 2508MHz / 218fsb
Windows 2000 - MingW32 GCC 3.4.1 PGO march=athlon-xp
Nodes: 550778
Best time: 1161 ms
Nodes per second: 474399 (Score: 1.911)
Athlon 64 3400+ 2.42GHz / 220fsb
Windows XP - MingW32 GCC 3.4.1 PGO march=athlon-xp
Nodes: 550778
Best time: 1187 ms
Nodes per second: 464008 (Score 1.908)
Athlon XP 2508MHz / 218fsb
Linux - GCC 3.4.1 PGO march=athlon-xp
Nodes: 550778
Best time: 1211 ms
Nodes per second: 454812 (Scores: 1.870)
Athlon 64 3500+ 2.2GHz / 200fsb
Windows 2000 - MingW32 GCC 3.4.1 PGO march=athlon-xp
Nodes: 550778
Best time: 1281 ms
Nodes per second: 429959 (Score 1.769)
Athlon XP 2083MHz / 166fsb (Barton core)
Windows 2000 - MingW32 GCC 3.4.1 PGO march=athlon-xp
Nodes: 550778
Best time: 1392 ms
Nodes per second: 395673 (Score: 1.627)
Athlon 64 3200+ 2.2GHz / 200fsb
Windows 2000 - MingW32 GCC 3.4.1 PGO march=athlon-xp
Nodes: 550778
Best time: 1438 ms
Nodes per second: 383018 (Score 1.575)
Athlon Tbird 1500MHz / 187fsb
Windows 2000 - MingW32 GCC 3.4.1 PGO march=athlon-xp
Nodes: 550778
Best time: 1932 ms
Nodes per second: 285081 (Score 1.172)
Athlon Tbird 1300MHz / 100fsb
Windows 2000 - MingW32 GCC 3.4.1 PGO march=athlon-xp
Nodes: 550778
Best time: 2263 ms
Nodes per second: 243384 (Score: 1.001)
Doesn't surprise me Anand is pulling this crap again. Tomshardware and HardOCP has been doing it for a *LONG* time. Anyone rememeber the faked Celeron 300a@600 at hardocp? How about the P4-2.53 beating the Athlon XP overclocked to 2.5GHz? Ridiculous stuff.
Tom was probably the worst. I had an identical system to his (AMD sent it to me for Quake 3 DLL optimization), I ran the same tests Tom did and could not reproduce his results. I ran test after test.. finally I started disabling things in the bios. The *ONLY* time I could reproduce his results exactly was when I set my AGP speed to 1x, left my bus at 133MHz but ran the memory at 100MHz (also it being asynchronous makes the latency horrid), disabled bank interleave AND made the timings something like 3-4-4-x.
Also anyone remember the Seti@home test he ran, back when the HT processors came out? He faked the results to show a P4 in Seti@home with HT was 2x faster than a P4 without HT of the same MHz. Absolute festering retardation (seeing how Seti@home doesn't support SMP/SMT anyway, and if it did it wouldn't EVER be 2x faster).
ALSO Why didn't Anandtech use Crafty? This is a popular chess benchmark that is used in the SPEC suite and reflects the performance of most chess programs. It also has great 64bit support (2GHz Opteron in 64bit mode is faster than a Dual Xeon 2.8GHz). Doesn't surprise me Anand picked TSCP. It is the only chess engine I know of (Vincent Diepeveen knows I've (Aaron Gordon) been in the computer chess scene for many years now) that leans that much towards the P4. Of course, a P4 optimized binary and a completely unoptimize binary on the A64 is enough to skew the results to how Anand wanted.
Anyway, this stuff has been going on for years.. many reviews.. all kinds of stuff. This isn't even the tip of the iceberg.