by Grigoriy Gubankov
06/25/2003 | 06:08 PM
Yesterday’s critical article on this web-site about the G5 benchmarks from Apple received so much attention from hardware community that today Greg Joswiak, Vice President of Hardware Product Marketing for Apple Computer gave a phone interview in which he defended his company’s performance claims for the G5 processor.
<%BANNER[article]%>At first, he said that SSE2 instructions were turned on during compilation of SPECCPU2000 tests. Having read the test report from VeriTest and GCC web-site, we have to say that yesterday’s claim about absence of SSE2 support was not correct. This claim came from misunderstanding of GCC optimization flags.
Mr. Joswiak added that the Hyper-Threading feature of Intel Pentium 4 processors was disabled in certain tests due to the fact that the Pentium 4 with Hyper-Threading enabled produced lower scores in these tests than «non-HyperThreaded» processor with the same frequency. Speaking about some modifications made to Apple systems, Mr. Joswiak said that memory read bypass option and software-based pre-fetching would be on by default on shipping systems. He is not sure only about modified malloc function that is faster, but less memory-efficient and, as VeriTest (a company which benchmarks hardware for Apple) says, “unsuitable for many users” (but perfectly suitable for benchmarking, I guess).
Answering the main question – why his company used GCC compiler for Intel-based systems, Mr. Joswiak said that this compiler was used because the test was intended to compare only hardware and not compilers, thus it would be logical to use the same compiler for all platforms. It appears that the reason for the performance drop in Pentium 4 scores in comparison to official SPEC scores is using GCC compiler, instead of Intel compiler. The main point here is that GCC in its present state is unable to produce vectorised SSE2 instructions, thus Intel Pentium 4 performance drops significantly.
I personally think that use of a single compiler for all platforms is doubtful decision, as in its G5 performance page Apple says that “Professional applications tested by Apple in June 2003 using preproduction Power Mac G5 units and, with the exception of HMMer, application software optimized for the PowerPC G5”. If you use software, optimized for G5, why don’t use the software, particularly the compiler, optimized for Intel Pentium 4?
Also, it is quite strange at first sight that there are no comparisons of G5 to AMD processors, dubbed Athlon XP and Opteron. I think the point here is that using GCC compiler instead of Intel’s will not result in big performance drop for AMD CPUs as they are less or not (in case of Athlon XP) dependant from SSE2 instructions. In addition, Opteron performs calculations using scalar SSE2 instructions much faster than Pentium 4 and even slightly faster than itself using vectorised SSE2 instructions (see here). Thus, it would be hard to set G5 in good light compared to AMD Opteron by switching to GCC compiler.
Editor’s Note: Apple has just made what it wanted to: it made a PC hardware web-site to post a news-story or two about Macs.