It refers to "performance portability", i.e. the ability to run a single piece of software at scale with good performance on a large variety of different system architectures - different CPU/GPU vendors, different architectures of different HPC systems, etc.
I'd expect a lot of the heavy lifting behind performance portability to come from templated frameworks like RAJA and Kokkos, or using OpenMP target offload. Well, at least until more support for native C++ and Fortran parallelism comes along.
I like the ICF implosion progression images on the hallway walls.
Those pictures on the wall.........implosion events that are computer modeled?
When they mention portability are they referring to apptainer?
It refers to "performance portability", i.e. the ability to run a single piece of software at scale with good performance on a large variety of different system architectures - different CPU/GPU vendors, different architectures of different HPC systems, etc.
I'd expect a lot of the heavy lifting behind performance portability to come from templated frameworks like RAJA and Kokkos, or using OpenMP target offload. Well, at least until more support for native C++ and Fortran parallelism comes along.
@@wileamyp Usually its a mix of MPI for inter-node parallelism, and any of your mentions (and more) for on-node parallelism.
But does it run Doom? 😂
Of course, someone's gonna have to port it to use ROCm first
VIVA LA RAZA
Next leap in super computing will be quantum computing