From Theory to Practice: Unleashing the potential of GPUs at the TREX Hackathon III

From Theory to Practice: Unleashing the potential of GPUs at the TREX Hackathon III

09 May 2023

On 06-08 March 2023, the third TREX Hackathon was held in Bologna at the premises of TREX partner CINECA. The Hackathon engaged 11 TREX code developers who worked over three days to port their codes to run on GPUs, or optimise their applications that already run on GPUs.  

The hackathon provided a unique opportunity for participants to learn more about GPUs, develop performance enhancements, and improve the portability of TREX flagship codes to new architectures. The Hackathon was structured in a mix of theoretical lessons and hands-on sessions supervised by experts Matt Bettencourt (NVIDIA) Giacomo Rossi (Intel).

Theoretical Sessions:
One and a half day was dedicated to theoretical sessions on general introduction to GPU architectures and GPU programming. A specific session was dedicated to the Marconi100 machine which was used by the participants to run the hands-on sessions on the second and third day. Experts from NVIDIA and Intel  gave an introduction to OpenACC and OpenMP with relevant examples and helpfully answered all the participants' questions.

Practical Sessions:
On the second day, a guided tour was organised to the Marconi100 supercomputer used for the hackathon, which had NVIDIA V100 cards installed, and  to the other CINECA machines, including G100, Marconi and DGX, which were equipped with NVIDIA A100 cards. The remaining time was focused on compiling the GPU version of qmckl on M100 and interfacing this GPU version of qmck library with TREX codes TurboRVB and Champ.

Main results of the hands-on session:

  • The device pointer version of qmckl_gpu (AO +MO) was run for the first time on Marconi's NVIDIA V100.
  • The Woodbury kernel for GPU has been fixed.
  • Progress has been made on interfacing TurboRVB and CHAMP to qmckl_gpu.

The event provided a unique opportunity for participants to learn more about GPUs and develop performance improvements, as well as improve the portability of TREX flagship codes to new architectures. The event was very interactive, and the participation of experts from NVIDIA and Intel allowed attendees to ask questions and get helpful answers.

Have a look at the materials from the Hackathon here.