Is Intel Throttling MKL Performance on AMD?
Added 2020-09-02 04:59:17 +0000 UTC
Comments
Ta. Any info on what may lie behind the AMD-specific changes? I could see that making sense if the intent is to hive off compiler suite development from INTEL, at which point AMD hopping onto the funding train would make sense.
Robert Bernecky
2020-09-02 18:54:22 +0000 UTCActually the ftc settlement required a bit more than that but those requirements expire next month. This is the polite video. The next video won't be as polite if things continue on this path.
Level1 Techs
2020-09-02 18:47:57 +0000 UTC(a) portable across new INTEL CPUs, (b) non-vendor-specific, which would produce better code for AMD CPUs, and (c) easier to test than CPUID-specific tests. This was viewed (and still is, by many of us) as a deliberate act by INTEL to give their CPUs better performance then AMD. The result of the case was, as I recall, a warning/notice on the INTEL compiler suite to the effect of "Your mileage may vary", without any other changes to the compiler to enhance its performance across vendors. This is one reason I never even try using INTEL compilers any more for my research. This nothing to do with the compiler development group at INTEL; it is purely a marketing ploy, intended to increase their market share. As for the new work with AMD-specific changes, I have no idea what they are intended to do, or whether INTEL has cleaned up their act on the software end of things. If the latter, I welcome it. That, of course, assumes that they get rid of the remaining CPUID checks, soon.
Robert Bernecky
2020-09-02 18:44:08 +0000 UTC
