You may have missed this writeup about UltraSparc T2 and Tigerton Tests which looked at low-level memory access measurements: http://blogs.sun.com/psa/entry/ultrasparc_t2_sun
A quote from a Sun employee I like... "You can only compute as fast as you can move data"
|
|
|
|
Good data but invalid conclusion. You can't compare results using different thread counts, which is what they did. I don't know about T2, but Tigerton cores can run more than 1 thread.
The data does show that the T2 latency has a large jump at 16 threads while the Tigerton system has not changed significantly. Thus, the T2 is already showing signs of stress at 16, while the Tigerton is not.
The experiment should be run out to 64, 128 or more threads and data compared at equal thread counts.
Posted by Xray on January 09, 2008 at 07:01 AM PST #
40 amd benchmarks blackbox blade blogwatch cmt cryptography e25k e6900 eco fire fluent hp hpc ibm intel itanium itanium2 java jvm linpack linux lotus m8000 m9000 mcad mysql nehalem networking niagara opensolaris opensource opensparc opteron oracle p595 performance power power5+ power6 powerflow record sap sap-sd scaling solaris sparc sparc64 spc-2 spec specapc specfp_rate2006 specint_rate2000 specint_rate2006 specjappserver specjappserver2004 specjbb specjbb2005 specomp specomp2001 specompl2001 specompm2001 specpower storage storagetek studio studio11 sun sunstorage7000 sunw suse swap t1 t2 t2000 t5120 t5220 t5440 tpc-c tpc-e tpc-h ultra ultrasparc us-iv+ utilization virtualization watt/performance wattage watts web2.0 woodcrest x4100 x4200 x4500 x4600 x6250 x64 xeon zfs
This theme brought to you by the Sun Green Team.
Go Green!
Good data but invalid conclusion. You can't compa...