 |
 |
On Dec 15, 2003, at 6:33 PM, Aleksandar Donev wrote: > Hello, > > My department is getting a new cluster and they are thinking about > what kinds > of processors to use. Among the options is G5. > > I would appreciate any info about Fortran 95 on G5 platforms. As far > as I > know, Absoft, NAG and now IBM (still beta) support this chip on OS X. > Experiences as well as any information on web-pages to look at for info > (pricing, requirements, etc.) would be helpful. > "We write quantum chemistry software (ADF) for several platforms, including OS X. We recently had access to a Dual 2Ghz G5, and did some testing. Here is what I posted to the scitech list at Apple: G4 466Mhz, 512Mb RAM, 1 processor Absoft Compiler (-O) CPU= 9394.72 Elapsed= 11076.76 XLF Compiler (-O3 -qhot -qtune=g5 -qarch=g5) CPU= 6003.03 Elapsed= 6284.52 G4 1Ghz, 2 processor Absoft compiler, 2 processor with PVM CPU= 2570.18 Elapsed= 2605.85 G5 2GHz, 512Mb RAM, 2 processor Absoft Compiler (-O) CPU= 1963.01 Elapsed= 1968.87 XLF Compiler (-O3 -qhot -qtune=g5 -qarch=g5) CPU= 1291.18 Elapsed= 1294.25 XLF Compiler (same flags), 2 processors, PVM CPU= 739.54 Elapsed= 800.63 Dual Xeon PIV 2.4 GHz, 1 CPU PGF90 Compiler CPU= 2091.60 Elapsed= 2193.39 IFC Intel compiler (-O3) CPU= 1143.88 Elapsed= 1272.71 Dual-CPU Pentium IV cluster, 2.4 GHz IFC compiler, 1 processor CPU= 1273.81 Elapsed= ??? IFC compiler, 2 processor CPU= 784.35 Elapsed= ??? 1.5 GHz Madison itanium2 IFC compiler, 1 processor CPU= 840.20 Elapsed= 886.14 IFC compiler, 2 processor CPU= 432.58 Elapsed= 460.48 1.5 GHz Itanium2 HP Compiler, 1 processor CPU= 586.78 Elapsed= 604.61 My conclusions: The XLF compiler produces code that is around 30-40% faster than Absoft, at least for ADF. I am a bit disappointed by the G5's performance with ADF. The times for the G5 are approximately in proportion to those for my old G4, once clock speed has been accounted for. I was hoping that architecture improvements in the G5, such as two FPUs, and that phenomenal frontside bus, would be seen in the benchmarks, but ADF doesn't seem to gain much from them. The times for a 2GHz G5 are very similar to those of a 2.4GHz Pentium, when using the Intel compiler on the Pentium (the portland group compiler produces considerably slower code). The G5 does seem to scale a bit better when run on two processors than the Pentium. I have no benchmarks for a 3.2GHz Pentium, but assuming that performance scales with clock speed, Intel still has the upper hand for running ADF. The difference for other programs seems to be less than for ADF though. Lastly, the new Itanium2 is an efficient and fast chip, but also more expensive. The HP compiler seems to be uncharacteristically good ;-)" Since writing these comments, I now think I know why the G5 is not great for ADF. ADF is old code, and is basically written as a bit dot product. The G5 has 2 FPUs, and two load/store units. A dot product requires two load/stores per multiply, and so basically the G5 is only using 1 FPU. This explains why it is not really any faster than the G4, when for many codes it is. So, if you code is well optimized, I think you will find the G5 to be very competitive with the fastest pentiums of the day, even with half the clock speed. Drew McCormack Theoretical Chemistry Free University, Amsterdam