GPU Servers

Waiting for parts.

OpenCL

No news.

Highly Ionizing Particle

Current version:

• Single warp streams neighboring pads in cacheline
• rocprof: 65% Memory Unit utilization (throughput: 130 GB/s ???)

New version:

• 576 threads to read full row (140 pads * 8 timebins)
• Most rows smaller than 140 pads -> wasted threads
• Memory Unit utilization drops to 36% 
• Slower by factor 1.57 (= 65 / 36)