Basic Roofline Model for Arria 10

In the HPC community a rough model based on 2 ceilings : the peak flops and the bandwidth is used as

a framework to evaluate architectures and algorithms (Sam Williams,D.Patterson The Roofline Model).

This one is for the Arria 10 FPGA, where the measured PCIe gen3 b/w is around 6GB/s and the peak flop/s is around 1.5 Tf/s.

A simple vector addition c = a + b[i] makes 1 flops every 12 bytes transferred therefore the

Arithmetic Intensity is 1/12 and the limit performance is 6*1/12 = 0.5 gflop/s.

http://www.alteraforum.com/forum/attachment.php?attachmentid=12856&stc=1