Current State

[kloppstock]

Currently, the following things are implemented and working:

• energy conversion algorithm
• photon conversion algorithm
• clustering algorithm
• summing up energy frames
• pedestal calibration
• online pedestal updates
• maximum value extraction
• pixel masking
• multi detector processing

All configurations were benchmarked, however the following yielded in high than expected runtimes:

• comparison of different cluster sizes
• comparison of different cluster counts
• multi detector benchmarks due to the problems mentioned below

The following general problems were found:

• high UPI usage when copying from GPU memory to RAM
• a performance limitation on GPU when approaching around 50 GB/s throughput
• non realtime capability of many CPU configurations

Which further steps were taken?

• The CPU and GPU implementation were thoroughly benchmarked and evaluated to be near the optimum when running on a real world data set
• all reasonable optimisations were applied for both GPUs and CPUs

Further documentation is available in the manuscript_paper.zip and jungfrau-photoncounter-report.zip in the cod folder.