Per Core Performance, the MHz really doesn’t tell you much now

Ah, the good old days.... It was normal to have a discussion with a friend or coworker member about something like, "We just bought a 1.2 GHz Pentium III server, it runs circles around that 500 MHz system we bought a few years back."  Everyone nods in approval, all rightly assuming that of course bigger is better and frequency directly relates to performance.  Of course now things are more complex with multi-core, multi-threads, differing architectures (Power, SPARC, Xeon, Opteron).  Is a dual-core at Power6 4.7 GHz faster than a Xeon at 3 GHz? Is a 1.4 GHz processor with 8 threads/core better than a 2.8 GHz quad-core with 2 threads per core?  Tough to know off the top of your head these days.  One thing is clear, the Intel Xeon processor 5500 series is in the lead of performance per processor (regardless of the frequency of processors available today). 

In comparing the Intel Xeon processor 5500 series (Nehalem) architecture vs. what's available from IBM, Sun, and AMD today, you see a wide variety of cpu offerings with dramatically differing specs.  However, when you take a look at all these systems with a common number of cores, you can see the differences in per core performance on the industry standard benchmark SPECint_rate_base2006


# of cpus

Total Cores

Total Threads


SPECint_rate_base2006 Performance

Intel Xeon X5570




2.93 GHz


AMD Opteron 2393SE




3.1 GHz


IBM Power6




4.7 GHz


Sun UltraSPARC T2




1.4 GHz


What a contrast!  Chip designers today have multiple choices to make to eek out the most performance in today's server systems.  What we see today is that the Intel Xeon processor 5500 series balances all of these quite well.  Whereas others have much higher frequencies, it doesn't necessarily translate into more performance, while others have gone with a larger number of threads, but have low performance per thread.  Even processors that have similar specs have performance that is quite different.  Of course this is only one benchmark, however if you look at others you will find similar differences.   

What this means for most IT buyers is it's more difficult to understand how all the whiz-bang features the marketers throw at you and how they translate into value for you.  My advice, really understand what kind of workloads are improtant to you and focus on the performance from industry standard workloads that best represent those.  Remember that bigger numbers on the spec sheet aren't always better when it comes to server performance.  Check your figures!

SPECint_rate_base2006 performance data reference:

Intel® Xeon® processor X5570 based platform details

Fujitsu PRIMERGY* TX300 S5 server platform with two Intel Xeon processors X5570 2.93GHz, 8MB L3 cache, 6.4GT/s QPI, 48 GB memory (6x8 GB PC3-10600R, 2 rank, CL9-9-9, ECC), SUSE Linux Enterprise Server 10 SP2 x86_64 Kernel, Intel C++ Compiler for Linux32 and Linux64 version 11.0 build 20010131. SPECint_rate_base2006 score 240,

AMD Opteron 2393SE based platform details

Supermicro A+ Server 1021M-UR+B, AMD Opteron 2393 SE 3.1 GHz, 6MB L3 cache, 32 GB memory (8x4 GB DDR2-800, CL5, Reg, Dual-rank), SuSE Enterprise Server 10 (x86_64) SP1, Kernel, PGI Server Complete Version 7.2, PathScale Compiler Suite Version 3.2, SPECint_rate_base2006 score 122,

IBM Power6 based platform details

IBM system p570 (4.7 GHz, 8 core), 32MB L3 cache, 64 GB memory (32x2 GB)DDR2 667 MHz, IBM AIX5L V5.3, XL C/C++ Enterprise Edition Version 9.0 for AIX, SPECint_rate_base2006 score 206,

Sun UltraSPARC T2 plus based platform details

Sun SPARC Enterprise T5120, Sun UltraSPARC T2 1.417 GHz, 4MB L2 cache, 64 GB memory (16x4 GB), Solaris 10 8/07 (build s10s_u4wos_12b), Sun Studio 12 (patch build 2007/08/30), SPECint_rate_base2006 score 73.0,