What is Good Performance?

Two fundamental requirements:

1. Must operate the GPU in the throughput-limited regime (not sensitive to Amdahl effects, D2H/H2D transfers, launch overheads, etc.)
2. Must attain high utilizatioin of the GPUs compute and/or bandwidth capabilities

Roofline Model is a throughput-oriented performance model

• Understand performance differences between Architectures, Programming Models, implementations, etc.
  • Why do some Architectures/Implementations move more data than others?
  • Why do some compilers outperform others?
• Predict performance on future machines / architectures
  • Set realistic performance expectations
  • Drive for HW/SW Co-Design
• Identify performance bottlenecks & motivate software optimizations
• Determine when we’re done optimizing code
  • Assess performance relative to machine capabilities
  • Track progress towards optimality
  • Motivate need for algorithmic changes