Dr David Carreto Fidalgo | Scientific AI workloads at the MPCDF
Institutskolloquium (intern)
- Datum: 15.04.2024
- Uhrzeit: 15:00 - 16:00
- Vortragende(r): Dr David Carreto Fidalgo
- Department of Neurophysics, MPI CBS, Leipzig, Germany & Max Planck Computing and Data Facility, Garching, Germany
- Ort: MPI für Kognitions- und Neurowissenschaften
-
Raum:
Lecture Hall (C101) + Zoom Meeting (hybrid mode)
https://zoom.us/j/94651679346?pwd=VWErd1hTc0ZnanBuQjYyWXF6Ti9TUT09 Meeting ID 946 5167 9346 Passcode 361703 - Gastgeber: Abteilung Neurophysik
- Kontakt: peter@cbs.mpg.de
In today's era of computational science, approaches powered by artificial intelligence (AI) have found their way into nearly every scientific domain. From natural sciences to the humanities, especially the emergence of large language models (LLMs) and their general purpose capabilities have opened new ways of processing unstructured data in a scientific context. However, training state-of-the-art AI models, even running them for inference, requires substantial computational resources. High-Performance Computing (HPC) systems, like the ones at the Max Planck Computing and Data Facility (MPCDF), are fitting solutions for executing these demanding AI workloads, and their integration into HPC workflows is increasingly sought after by the scientific community. In this talk we showcase two projects in which such workloads are successfully handled at the MPCDF in close collaboration with the Neurophysics group at the CBS. These projects use AI models to reconstruct and automatically segment high-resolution MRI images. Additionally, we will present a couple of projects revolving around LLMs at the MPCDF and highlight their computational challenges and solutions for HPC systems.
In today's era of computational science, approaches powered by
artificial intelligence (AI) have found their way into nearly every
scientific domain. From natural sciences to the humanities, especially
the emergence of large language models (LLMs) and their general purpose
capabilities have opened new ways of processing unstructured data in a
scientific context. However, training state-of-the-art AI models, even
running them for inference, requires substantial computational
resources. High-Performance Computing (HPC) systems, like the ones at
the Max Planck Computing and Data Facility (MPCDF), are fitting
solutions for executing these demanding AI workloads, and their
integration into HPC workflows is increasingly sought after by the
scientific community.
In this talk we showcase two projects in which such workloads are
successfully handled at the MPCDF in close collaboration with the
Neurophysics group at the CBS. These projects use AI models to
reconstruct and automatically segment high-resolution MRI images.
Additionally, we will present a couple of projects revolving around LLMs
at the MPCDF and highlight their computational challenges and solutions
for HPC systems.