Intel(R) MKL 11.0 introduces new functionality allowing users to balance the need for reproducible results with performance. The webinar recording and presentation linked below discusses the mechanisms that cause variability in floating point results, the new controls to limit these, and the performance trade-offs involved.
MKL CNR
Conditional Numerical Reproducibility (CNR) in Intel MKL 11.0
New functionality in Intel MKL 11.0 now allows you to balance performance with reproducible results by allowing greater flexibility in code path choice and by ensuring that algorithms are deterministic. To learn more about Conditional Numerical Reproducibility (CNR) see the following resources:
Introduction to Conditional Numerical Reproducibility (CNR)
Intel® MKL 11.0 introduces a feature called Conditional Numerical Reproducibility (CNR) which provides functions for obtaining reproducible floating-point results when calling library functions from their application. When using these new features, Intel MKL functions are designed to return the same floating-point results from run-to-run, subject to the following limitations:
Intel® Math Kernel Library 11.0
This document contains details about the Intel® Math Kernel Library 11.0 (Intel® MKL 11.0) new features.
Webinar: Getting Reproducible Results with Intel® MKL 11.0 beta
A technical talk on the Condition Numerical Reproducibility (CNR) feature in Intel® MKL 11.0
