User Guide

Contents

GPU Roofline Accuracy Presets

For each perspective, you can choose collection accuracy level. Each accuracy level is a preset of analysis stages and properties with a corresponding accuracy value. The higher accuracy value you choose, the higher runtime overhead is added to the application you run with
Intel® Advisor
.
The following accuracy levels are available:
Accuracy
Runtime Overhead
Analyses and Properties
Results
Low
5 - 10x longer
Survey with GPU profiling + Characterization (FLOP)
GPU Roofline chart for all memory levels reported for your application kernels executed on GPU:
  • Plotted for all memory levels (CARM, L3, SLM, GTI)
  • Kernel is represented by multiple dots for different memory levels
  • Metrics related only to the loop/function itself excluding data originating in other loops/functions called by it (self data)
Basic CPU Roofline report reported for your application loops/functions executed on CPU
Medium
10 - 50x longer
Survey with GPU profiling + Characterization (Trip Counts, FLOP, call stacks for CPU, CPU cache simulation)
GPU Roofline chart for all memory levels reported for your application kernels executed on GPU.
CPU Roofline chart for all memory levels with total data reported for your application loops/functions executed on CPU.
You can choose custom accuracy and set a custom perspective flow for your application. For more information, see Customize
GPU Roofline Insights
Perspective
.
There is a variety of techniques available to minimize data collection, result size, and execution overhead. Check Minimize Analysis Overhead .

Product and Performance Information

1

Performance varies by use, configuration and other factors. Learn more at www.Intel.com/PerformanceIndex.