Eliane Briand

@ebriand.bsky.social

51 Followers 77 Following

doin' simulations (MPI-NAT Goettingen) https://eliane.br.iand.fr/

First slide for a "Why we do not use Kokkos" deck, from Sandia Lab (famously, maker of Kokkos)
"Get ready to learn chinese buddy"-type meme, except, it's NVidia CUTLASS not Chinese
CUDA code for a similar kind of extended precision summing code, here a fast2sum based cascading scheme. This takes a cascading level as template argument, with 1 being normal double, 2 using two double, one for running sum, another for the tally of the errors, and 3 and beyond holding the errors of the error sums of the level before.
C/C++ source code for a Kahan sum accumulator using __float 128 data type, the fabled quadmath floating point type. Kahan sum allows the additions of floating point number using two elements, one for the running sum, one for the running floating point addition error, providing an extended precision way of summing large arrays of disparate numbers, here almost equal to an "octuple" precision floating point number
Graph showing the number of deposited experimental structure on the PDB. It is steadily rising without clear discontinuity since 2020-2021.
3d render of a ceramic cup full of very much not to scale water molecule (space filling sphere), nearby a protein in ribbon representation, over a science paper as table.