Performance engineering: being friendly to your hardware
summary
Shows measurable hardware effects: memory goodput sits near ~10% of interface time, banking and page mapping cost matters, HBM helps only with parallel requests, and enhanced instructions like enhanced rep movs let the core choose microcode paths. Branch prediction reaches ~90% but misprediction flushes cost cycles. Vectorization boosts throughput (example claim AVX-512 16× speedup); practical profilers named are VTune and MCA. Slides and code patterns are given and concrete microbenchmarks and mappings are shown.
AI-written summary · about
chapters (7)
- 0:00Hardware-friendly perf engineering
- 2:07How memory systems work
- 15:20Processor cores basics
- 37:49Branch prediction examples
- 43:15Optimizing memory copies
- 51:43Why micro-optimizations matter
- 1:00:37Q&A profiling tools
related talks
01h 05mPerformance Optimization in Software Development - Being Friendly to Your Hardware
01h 04mDesign for Performance
01h 37mAn Overview of Program Optimization Techniques
01h 02mSqueezing the Hardware to Make Performance Juice
58mOptimization Tips - Mo' Hustle Mo' Problems
51mDesigning for Performance
50mDesigning for Performance
37mHigh Performance Computing: The Modern Approach