- Local: Andrea Sciabà, Servesh Muralidharan
- Remote: Andrea Valassi, Michele Michelotto, Stefan Roiser, Chris Hollowell, Mattieu Puel, Felice Pantaleo, Luca Atzori, Laurent Duflot, Shigeki Misawa
We went first through the slides. Some highlights from the discussion:
- Felice explains that professional GPUs are much better than gaming cards because of the FP64 units (32x more in the former for Nvidia)
- OpenCL was promising, but its popularity is falling considerably
- Somebody in CMS uses TensorFlow in a Higgs analysis to run on any kind of GPU
- At a recent Hackaton, somebody showed some work on beam dynamics using OpenCL; will ask for performance plots
- Professional GPUs are much better also for the cooling; gaming GPUs would not have any support from Nvidia
- Professional GPUs have also much more memory; this is not very important for CMS reconstruction now, but it is important for ML applications. In addition, memory bandwidth is much higher and there is also error control
- Felice adds that from the programmability point of view, in the medium/long term, CMS will move to a shared memory paradigm between CPU and GPU, where GPU memory will act as a cache. The PCIe bus could become the bottleneck.
- Servesh asks abut OpenMP: Felice explains that it's a scheduler and the experiment framework is already a scheduler
We then went through the document. Highlights:
- About interconnect technology, as it's mainly related to CPUs, it falls inside the domain of this subgroup
ACTIONS
- All: continue writing the document and the slides
Next meeting is scheduled for February 22 at 16:00.
There are minutes attached to this event.
Show them.
