Getting ready for MI300
In early 2024, DiRAC collaborated with AMD to present two events aimed at preparing our users for the new MI300A APU.
AMD GPU All–User Training Event
Held remotely over three days at the end of January, this series of lectures enabled researchers at any experience-level to gain the knowledge and skills needed to leverage the capabilities of the MI300A APU and acted as a pre-training course for hackathon attendees. Presented by trainers from AMD, the course looked at:
- 3D Chiplet Technology: Combining CPU and GPU chiplets on the same die
- On-Chip High-Speed Interconnect: Facilitating efficient communication between CPU and GPU
- Unified Shared Memory Model: Providing a seamless and unified approach to memory management
- Single memory address space: Simplifying memory access as different processes refer to the same byte of data
The lectures were recorded and will be available on our website later in the year.
AMD GPU Advanced-User Hackathon
A three day, in-person, hackathon was held in Durham two weeks after the all-user virtual training. The hackathon gave 6 teams from DiRAC’s core code groups and a team from UKAEA the opportunity to work on their own codes in the company of two AMD experts and members of the DiRAC RSE team. These mentors were joined virtually by members of AMD’s Compiler and Development teams to provide in-depth presentations and additional support during exercises.
Topics covered included:
- Programming Model for MI200 & MI300 Series
- OpenMP Offload on AMD GPUs
- Optimising HIP Applications
The hackathon work was only paused for several very successful pub visits, and a great networking dinner on the Wednesday night which gave an opportunity for more informal discussion. All agreed this was a super event, with participants and AMD trainers expressing a desire for a repeated AMD hackathon in the future.
Testimonials and comments
“…We found it extremely helpful to have AMD experts in the room which eventually went through our compiled code with us to identify flaws in the generated code such as extensive register pressure and excessive integer calculations…”
“…Their consulting helped us to work out a clear strategy how to tweak and tune our kernels for upcoming GPU generations.”
“…In collaboration with our partners from AMD, we used the hackathon to identify all ingredients from the standard libraries which we require, for our code, to be available on the accelerator, and we managed to port those that were not already there…”
“Support available was fantastic. The team were really skilled and knew how to best advise in all aspects of the process. I really enjoyed seeing other teams work and being able to share progress, ideas, and difficulties between teams.”