Skip to content
About
Main content
Model

The roofline model

Left of the ridge point, performance = intensity × bandwidth: only moving fewer bytes helps.

Performance
achievable performance (FLOPs per second)
Arithmetic intensity
FLOPs done per byte loaded from DRAM
Memory bandwidth
DRAM bandwidth, ~50–200 GB/s per socket
Compute peak
the flat compute roof: peak FLOPs per second of the cores

How it behaves

One input on the horizontal axis, the others held fixed.

Data table
Achievable performance by arithmetic intensity (illustrative: β = 100 GB/s, π = 1 TFLOP/s)
Arithmetic intensity I (FLOP/byte)Performance (FLOP/s)
0.011B
0.0111.12B
0.0131.26B
0.0141.41B
0.0161.58B
0.0181.78B
0.022B
0.0222.24B
0.0252.51B
0.0282.82B
0.0323.16B
0.0363.55B
0.043.98B
0.0454.47B
0.055.01B
0.0565.62B
0.0636.31B
0.0717.08B
0.0797.94B
0.0898.91B
0.110B
0.11211.2B
0.12612.6B
0.14114.1B
0.15815.8B
0.17817.8B
0.220B
0.22422.4B
0.25125.1B
0.28228.2B
0.31631.6B
0.35535.5B
0.39839.8B
0.44744.7B
0.50150.1B
0.56256.2B
0.63163.1B
0.70870.8B
0.79479.4B
0.89189.1B
1100B
1.12112B
1.26126B
1.41141B
1.58158B
1.78178B
2200B
2.24224B
2.51251B
2.82282B
3.16316B
3.55355B
3.98398B
4.47447B
5.01501B
5.62562B
6.31631B
7.08708B
7.94794B
8.91891B
101,000B
11.21,000B
12.61,000B
14.11,000B
15.81,000B
17.81,000B
201,000B
22.41,000B
25.11,000B
28.21,000B
31.61,000B
35.51,000B
39.81,000B
44.71,000B
50.11,000B
56.21,000B
63.11,000B
70.81,000B
79.41,000B
89.11,000B
1001,000B

Where it appears

Each mechanism states it in its own names. The general symbol each name stands for is shown with it.