SiSoftware Sandra (the System ANalyser, Diagnostic and Reporting Assistant) is an information & diagnostic utility. It should provide most of the information (including undocumented) you need to know about your hardware, software and other devices whether hardware or software. The tool is critically acclaimed and used in about every hardware review you'll read in PC IT journalism when CPU's, motherboards or memory sticks are involved. Warp2Search is an enlisted member of the SiSoftware Sandra beta program. Intensive testing on our own machines revealed that Sandra 9.44 and even the latest version Sandra MAX3 deliver inconsistent benchmark results on Intel Pentium 4 Hyperthreading capable CPU's using the CPU Arythmetic Benchmark. We have informed SiSoftware about our findings and continued to work with Adrian Silasi of SiSoftware to get to the bottom of the problem.
We performed tests using a 3.06Ghz HT P4 FSB533 CPU and a 3.0Ghz HT P4 FSB800 CPU. The first ran on an i850e chipset based motherboard, while the latter was set up on an i875 chipset motherboard. Windows XP Professional with Service Pack 1 served as the operating system. Hyperthreading remained enabled during all tests. Please note that this issue only applies to the CPU Arythmetic Benchmark. Our observation was, that benchmark results we could obtain with Sandra 9.44 were VERY inconsistent. We retested with Sandra MAX3 and the problem persisted, though to a lesser degree than with the previous version. The obvious problem lies in Dhrystone ALU results, which varied from ~8400 to ~9700 MIPS, which is totally out of the usual margin of errors. Whetstone FPU/iSSE2 results were more stable, but also varied in a range of about 300 MFLOPS. The instability in results also persisted if the CPU's were overclocked. Due to our good relationship to SiSoftware here in Europe, we were able to bring the findings directly to the desk of C. Adrian Silasi (B.Eng., M.Sc., MIEE SiSoftware). "It looks like the SSE2 version of Whetstone uses an old version of Intel's SSE2 math library that seems to corrupt the stack under some circumstances when running in SMP mode.", he said in our email correspondence. According to Mr. Silasi these circumstances are very rare and therefore up to 90% (possibly even higher than that) of the results achieved are representative of the real performance. A new math library from Intel has been received by SiSoftware that corrects this issue in calendar week 29, by which Sandra MAX3 already was out for about 2 weeks. Because of the new library which Sandra depends on, SiSoftware is forced to regress testing on all platforms (Celeron, P4, Xeon, Opteron (?), Banias) before it makes it into the new beta and hopefully in the next release. This process includes recalibration of the reference scores, we were told by Mr. Silasi. "We may drop this library from the next release since we cannot use this on our AMD64 version and Intel is unlikely to make a version for AMD64. We're looking at AMD's math library but we may be forced to write our own though this is non-trivial. Until this goes at least beta rather than alpha we cannot make any decisions anyway." - he continued to explain the circumstances involved in regress testing and recalibration. "I am not sure why we did not see this before - but we'd have gotten complaints if it were to happen. It may be, that the desktop P4 HT is more "susceptible" to these issues as our Xeon HT / Opteron / Banias boxes don't seem to have a problem with it." he concluded. Warp2Search continues to help SiSoftware to extinguish this issue and provide one of the best benchmark solutions to journalists, enthusiasts and SiSoftware customers. We will continue coverage on this story as it evolves.
We performed tests using a 3.06Ghz HT P4 FSB533 CPU and a 3.0Ghz HT P4 FSB800 CPU. The first ran on an i850e chipset based motherboard, while the latter was set up on an i875 chipset motherboard. Windows XP Professional with Service Pack 1 served as the operating system. Hyperthreading remained enabled during all tests. Please note that this issue only applies to the CPU Arythmetic Benchmark. Our observation was, that benchmark results we could obtain with Sandra 9.44 were VERY inconsistent. We retested with Sandra MAX3 and the problem persisted, though to a lesser degree than with the previous version. The obvious problem lies in Dhrystone ALU results, which varied from ~8400 to ~9700 MIPS, which is totally out of the usual margin of errors. Whetstone FPU/iSSE2 results were more stable, but also varied in a range of about 300 MFLOPS. The instability in results also persisted if the CPU's were overclocked. Due to our good relationship to SiSoftware here in Europe, we were able to bring the findings directly to the desk of C. Adrian Silasi (B.Eng., M.Sc., MIEE SiSoftware). "It looks like the SSE2 version of Whetstone uses an old version of Intel's SSE2 math library that seems to corrupt the stack under some circumstances when running in SMP mode.", he said in our email correspondence. According to Mr. Silasi these circumstances are very rare and therefore up to 90% (possibly even higher than that) of the results achieved are representative of the real performance. A new math library from Intel has been received by SiSoftware that corrects this issue in calendar week 29, by which Sandra MAX3 already was out for about 2 weeks. Because of the new library which Sandra depends on, SiSoftware is forced to regress testing on all platforms (Celeron, P4, Xeon, Opteron (?), Banias) before it makes it into the new beta and hopefully in the next release. This process includes recalibration of the reference scores, we were told by Mr. Silasi. "We may drop this library from the next release since we cannot use this on our AMD64 version and Intel is unlikely to make a version for AMD64. We're looking at AMD's math library but we may be forced to write our own though this is non-trivial. Until this goes at least beta rather than alpha we cannot make any decisions anyway." - he continued to explain the circumstances involved in regress testing and recalibration. "I am not sure why we did not see this before - but we'd have gotten complaints if it were to happen. It may be, that the desktop P4 HT is more "susceptible" to these issues as our Xeon HT / Opteron / Banias boxes don't seem to have a problem with it." he concluded. Warp2Search continues to help SiSoftware to extinguish this issue and provide one of the best benchmark solutions to journalists, enthusiasts and SiSoftware customers. We will continue coverage on this story as it evolves.
