IEEE floating point does not adhere to associative and distributive laws of arithmetic. This creates rounding noise due to ordering effects of concurrent operations.
Posits are a tapered floating point system that restores associativity and distributive arithmetic without rounding noise. We have build a High-Performance Reproducible linear algebra environment using posit arithmetic, but for DMM we need to augment MPI reductions to use posit arithmetic.
We are looking for a research collaboration to add a posit-enable reduction instruction set to OpenMPI to create a fully scalable high-performance reproducible linear algebra solution for arbitrary concurrency environments.
The benefit of this effort would be that research collaborations between teams that have very different cluster configurations becomes trivial, because computational results are independent of the concurrency dynamics of the cluster.
New feature request
Details of the problem
IEEE floating point does not adhere to associative and distributive laws of arithmetic. This creates rounding noise due to ordering effects of concurrent operations.
Posits are a tapered floating point system that restores associativity and distributive arithmetic without rounding noise. We have build a High-Performance Reproducible linear algebra environment using posit arithmetic, but for DMM we need to augment MPI reductions to use posit arithmetic.
We are looking for a research collaboration to add a posit-enable reduction instruction set to OpenMPI to create a fully scalable high-performance reproducible linear algebra solution for arbitrary concurrency environments.
The benefit of this effort would be that research collaborations between teams that have very different cluster configurations becomes trivial, because computational results are independent of the concurrency dynamics of the cluster.