10-Year Anniversary Workshop of NVIDIA Application Lab at Jülich
At ISC 2012, JSC and NVIDIA signed a collaboration agreement to enable and optimize scientific applications for upcoming GPU-based systems; the NVIDIA Application Lab was then founded at Jülich. Almost 10 years and about 5000 NVIDIA GPUs later, a workshop was held on 21 and 22 June 2022 to celebrate this anniversary and look back on the last 10 years of the lab.
A number of application developers with whom the Application Lab has collaborated over the years presented past developments, recent challenges, and future plans. In addition, two hardware-related talks were held, focusing on JSC’s GPU installations (most importantly the JUWELS Booster) and NVIDIA’s new Hopper H100 GPU processor.
The application presentations included, for example, a report about the GPU activities with JUQCS, the Jülich Universal Quantum Computer Simulator. This application is accelerated with CUDA Fortran and is able to utilize more than 2000 GPUs of the JUWELS Booster with excellent scaling and complex communication patterns. Another example is MPTRAC, an application developed in the SDL Climate Science. The application is accelerated with OpenACC and developers are currently investigating methods to deal with the large input data sets. In a talk about the fluid dynamics simulation D2Q37, a developer from the University of Ferrara/INFN Ferrara presented recent advanced transporting GPU optimizations to FPGA-based systems. D2Q37 had already been presented during the first Application Lab Workshop 10 years ago. At that time, the GPU acceleration had just started.
Lab colleagues from NVIDIA presented the new H100 GPU in detail, which introduces features such as multiprocessor clusters. Such a cluster has as many multiprocessors as an entire K20X GPU had 10 years ago – now with much more performance per multiprocessor. More information on the workshop can be found in the blog post.
Since the founding of the lab, more than 50 GPU-related activities and training courses were hosted (e.g. programming courses, hackathons, porting workshops, conference tutorials) and collaborations with various application developers were pursued. The intensive preparatory work of this lab and the experience gained in these 10 years provide an excellent basis for making applications fit for the upcoming GPU-based exascale system JUPITER.
Contact: Dr. Andreas Herten, firstname.lastname@example.org
from JSC News No. 290, 8 August 2022