PROCESSOR SPEED: Linpack, SPEC etc. This page is a compendium of approximate Linpack and SPEC numbers for some well-known processors on the market, mainly RISC, but also some of the leading vector processors. This page is intended and useful (only) for quick comparisons. The Linpack and SPEC numbers are sometimes taken from different sources and may not apply to a particular machine containing the processor referred to. Some processors perform markedly differently in different machines. However, the numbers are quoted from reliable sources and should represent good estimates. The primary aim of the following Table was to see whether there is any correlation between the SPEC and Linpack benchmarks. Later the Table grew to encompass most leading processor makes. The Table is not intended to be exhaustive nor is any entry guaranteed for correctness. Table I. Processor speeds. Sources: WWW pages, Linpack benchmark, etc.. The speeds are sometimes only approximate because a particular computer data were used. ============================================================================== Machine/Processor MFLOPS LP* MIPS SPECint92 SPECfp92 SPECint95 SPECfp95 ------------------------------------------------------------------------------ Alpha 5/300 21164 600 277 410 7.33 12.16 Alpha 5/433 21164 866 12 15 Alpha Station 600 5/500 1000 14.8 17.4 DEC 8100 5/440 880 588 DEC 500/500 1000 590 ------------------------------------------------------------------------------ HP 7100 735/99 (SPP-1000) 3.27 4.06 HP 715/75 31 86 80 127 HP 9000 J210 7200/120 4.37 7.60 735/125 PA-RISC 7150 4.04 4.61 9000/735 99 41 B 132L (7300/132) ~80? 5.9 6.2 PA-RISC 7200/100 136 215 PA-RISC 8000/160 C-160 640 421 10.4 16.3 PA-RISC 8000/180 C-180 720 480 11.8 18.7 Exemplar S-Class (1 proc.) 720 545 Exemplar V-Class (1 proc.) 800 613 PA-RISC 8200/220 ? 880 15.5 25 SPP-1000 (1 proc.) 200 Convex C3860 (1 proc.) 120 104 ------------------------------------------------------------------------------ SGI Power C. L R8000/75 300 250 *2 SGI Power C. L R8000/90 360 300 ** SGI Origin R10000/195 390 344 SGI Octane R10000/195 390 315 9.3 17 Indy MIPS R4000/100 PC ~25 19 Indigo 2 R4400/250 SC 87 176 165 Indy MIPS R4600/133 SC ~40 30 ~80 Indy MIPS R5000/150 SC ~60 50 O2 MIPS R5000/180 SC 4.8 5.4 Cray T94 (1 p., 2.2 ns) 1800 1603 Cray C916 (1 proc.) 1000 Cray C90 (1 p., 4.2 ns) 952 902 Cray J90 (J932, 1 proc.) 200 202 Cray T3D (DECchip 21064) ? 150 Cray T3E (Alpha? ?/450) 900 504 *1 Cray Y-MP/832 (1 p., 6 ns) 333 324 Cray X-MP/416 1 p., 8.5 ns 235 218 ------------------------------------------------------------------------------ NEC SX-4 (1 proc.) 2000 1944 ------------------------------------------------------------------------------ Hitachi S-3800/180 1/2 ns 8000 6431 ------------------------------------------------------------------------------ IBM RS/6000-R24/71.5 284 246 IBM ES/9000-711 VF 1/7 ns 563 422 ------------------------------------------------------------------------------ Fujitsu VPP2200/10 4 ns 1000 842 ------------------------------------------------------------------------------ ============================================================================== * If not otherwise noted, these numbers refer to the Linpack 1000x1000 "hard coded" test (as quoted by HP) or the "Best Effort" test (as quoted by Linpack); or matrix multiplication MFLOPS (400x400) as measured by myself on an Indy, using complib.sgimath (precompiled DGEMM). These numbers should be a reasonable floating point performance measure. ** Linpack gets 308 so my matrix mupltiplication test is very close. *1 Data by Cray. *2 Linpack gets 261. There are several interesting things to observe. One is a low efficiency (LP vs. MFLOPS) for some processors/machines. The other is the small correlation between SPECfp and Linpack across different processors/machines, as shown by the following Table: Table II. Ratios SPECfp95 vs. theoretical and realistic (Linpack etc.) MFLOPS from Table I. ========================================================== Processor SPECfp95 : MFLOPS SPECfp95 : LP ---------------------------------------------------------- Alpha 5/500 0.017 0.029 PA-RISC 8000/180 0.026 0.039 R10000/195 0.044 0.054 ========================================================== The following Table shows the varied floating point efficiency of processors in different machines. Table III. Floating point efficiency: realistic (Linpack etc.) MFLOPS from Table I vs. theoretical MFLOPS. ===================================== Processor LP : MFLOPS ------------------------------------- Alpha 5/500 0.59 PA-RISC 8000/180 0.67 R8000/90 0.83 R10000/195 0.88 ===================================== At the end, a small table from the Linpack report, where everybody can find the justification for his favorite processor (strictly joking, of course): ====================================================================================================================== Computer OS/Compiler N=100(Mflops) N=1000(Mflops) Peak(Mflops) ---------------------------------------------------------------------------------------------------------------------- DEC 500/500 (1 proc, 500 MHz) kf77 -inline=daxpy -ur=3 -fast -O5 -tune 235 590 1000 DEC 500/400 (1 proc, 400 MHz) kf77 -inline=daxpy -ur=3 -fast -O5 -tune 189 449 800 HP Exemplar S-Class(1 proc) SPP-UX 5.2+Oall +Oinline=daxpy 156 545 720 SGI POWER CHALLENGE (90 MHz,1 proc) -non_shared -OPT: IEEE_arithmetic=3: 126 308 360 Gateway 2000 G6-200 PentiumPro MS Fortran NT /G5 /Oxb2 62 200 Pentium Pro 200 Mhz Solaris 2.5 GNU F77 v0.5.5 38 200 DELL XMT5133 Pentium 133MHz PS 32 NT V 1.0 /G5 /Oxb2 14 Gateway 2000 P5-90(90 MHz Pentium) Windows NT /G5 /Oxb2 11 CECpx XL 560 Pentium 60 MHz 10.5 wfc386 /l=dos4g /ox 7.2 Gateway P5-60 (60 MHz Pentium) F77L-EM32 5.01 /4 /Z1 5.4 Gateway P5-60 (60 MHz Pentium) 77/32/mf/d1/warn/5/fp5/ot 5.3 Gateway Pentium (66HHz) Lahey F77, 4.00 1.2 ====================================================================================================================== Note: 100x100 is a fixed benchmark, while 1000x1000 is "best effort". (One wonders: why are there no best efforts for some processors. Memory?) Clearly, 100x100 is a prototype of "my favorite benchmark" ("useless", in English). Compiled by R. Krivec May 1997