Profile cover photo
Profile photo
Karl Rupp
Computational semiconductor scientist.
Computational semiconductor scientist.

Karl's posts

Post has attachment
Inspired and kick-started by +Denis Demidov and his wife: My first home-made yoghurt! HPC^1 matters! :-)

^1 High Performance (yoghurt) Culture

Post has attachment
Hans Petter (Langtangen) will be missed :-(

Post has shared content
Our CSE codes get more and more complex. Let us not forget about the many sound tools and principles from software engineering to deal with the complexity. This workshop spanning many different institutions looked at how CSE can benefit from software engineering and vice versa.

This report details the challenges and opportunities discussed at the NITRD sponsored multi-agency workshop on Computational Science and Engineering Software Productivity and Sustainability (CSESSP) Challenges, held in Washington, D.C. USA on October 15-16, 2015. The workshop brought together 85 attendees from all branches of the U.S. federal government, industry, academia, and U.S. and international research laboratories to discuss growing concerns over the sustainability of our Computational Science and Engineering (CSE) software foundation, and the productivity of scientists and engineers who develop and use this software. Discussions focused on characterization of the challenges, and opportunities for improved productivity and sustainability going forward. This report comprises workshop and subsequent discussions including a summary of key opportunities for the CSE community going forward.

View Full Text:

Post has attachment
I'm baffled that Vincent is either
1) not aware of the use of log-scale for plotting data with exponential trend
2) intentionally using a linear scale to deceive readers.
Not to mention that he is comparing theoretical peak bandwidth on the GPU with actual STREAM bandwidth on KNL.

HPCwire, do you have any quality standards?

Post has attachment
My comparison of CPUs, GPUs, and MIC has received a 2016 update. Now with #Intel #KNL and #NVIDIA #Pascal data:

The comparison of 'cores' now considers a GPU streaming multiprocessor  (NVIDIA, compute unit for AMD) as a 'core'. This better reflects the SIMD-like lock-step execution of threads in a warp or wave front.

FLOPs per cycle is also pretty interesting:
Did you know that high-end CPUs today offer more FLOPs per clock cycle than high-end GPUs in 2008? I wasn't aware of it.

Post has attachment
I hope that public relation departments at major universities also get the message...

Post has attachment
My take on threads and how to deal with them in software libraries. The comparison with MPI towards the end is striking. I guess this overlaps with the issues +Matthew Knepley , +Barry Smith and +Jed Brown had with threading in PETSc.

Post has attachment

Post has attachment
It's good to see that the Software Sustainability Institute (UK) is very much in line with what +Wolfgang Bangerth and +Timo Heister already addressed in their SIAM News article 'Quo Vadis, Scientific Software?'.

Post has attachment
ViennaCL 1.7.0 is now available for download:

Highlights of this release:
* Fine-grained parallel incomplete LU factorization preconditioners
   (implementation inspired by )
* Fast sparse matrix-matrix products
  (implementation inspired by
* Fine-grained parallel algebraic multigrid preconditioners
  (extending to OpenCL and OpenMP)
Wait while more posts are being loaded