Hi, We are looking at how to manage the heterogeneity within our cluster.
The page https://docs.easybuild.io/controlling-compiler-optimization-flags/ refers to the following link https://easybuild.io/files/sciCORE-software-management_20150611.pdf The document referred to is a little light on detail, but includes the statement: we keep few identical software stacks. One software stack per cpu type (sandybridge, ivybridge..etc) We have the following CPUs Intel(R) Xeon(R) Gold 6130 CPU @ 2.10GHz Intel(R) Xeon(R) Silver 4210 CPU @ 2.20GHz which are both Skylake processors, but the 4210 has the extension AVX512_VNNI, whereas the 6130 does not. TensorFlow can make use of the extension if it is present, so if it compiled on the 4210 processor it will not run on the 6130 processor and exits with The TensorFlow library was compiled to use AVX512_VNNI instructions, but these aren't available on your machine. Has any one got any ideas about how one should deal with such differences below the level of the microarchitecture? Cheers, Loris -- Dr. Loris Bennett (Herr/Mr) ZEDAT, Freie Universität Berlin
