Tomorrow I'm starting a new role at NVIDIA, in the GPU communications team (the folks who bring you NCCL and NVSHMEM). For the past 4 years, I've been focused on building a standard software ecosystem around Grace Hopper and making it even easier to get scientific results on GPUs, particularly with Fortran. I also managed to do the (supposedly) impossible and standardize an MPI ABI. Now it's time to solve some different problems. GPU-centric computing is normative in AI and much of HPC now. GPUs need to orchestrate more and more of the I/O to scale workloads. My goal now is to make GPU communication easier to use and more effective for a wide range of workloads.
Welcome to the team!!
Sounds fun, Jeff! Now I know whom the PETSc team should bug about NVSHEM stuff. 😁 (We've added an NVSHMEM back-end in PETSc but haven't fully explored this -- it seems promising, though.)
Congrats Jeff!
A lot of our old quantum chemistry techniques are coming around full circle, just now with dynamic workers coming and going.
Congratulations, Jeff!
Nice!
Congrats Jeff!
#mpi 👍
Awesome for you, Jeff. Congratulations on this next chapter!
Managing Director, High Performance Computing Center at TTU; Co-Director, NSF Cloud and Autonomic Computing Center
8moGreat! Can we get better support for building the NVIDIA GPU ecosystem on a wider variety of distributions? I don’t like being limited to Ubuntu for things like the Jetson Orin Nanos that we use in training mini-clusters and instrumentation workloads.