Cosine similarity metric calculation

on low power heterogeneous

computing platform

Michał Karwatowski1,2, Sebastian Koryciak1,2,

Ernest Jamro1,2, Agnieszka Dąbrowska-Boruch1,2,

Kazimierz Wiatr1

1AGH University of Science and Technology, al. Mickiewicza 30, 30-059 Kraków,2ACK Cyfronet AGH, ul. Nawojki 11, 30-950 Kraków

KUKDM 11-13.03.2015 Zakopane

Agenda

FPGA based hardware accelerated

computing

Text similarity analysis

Search algorithm implementation

Results

Future work

2

FPGA based hardware accelerated

computing3

Text similarity analysis 4

Text comparison 5

Hardware 6

Zynq 7

Hardware architecture 8

Compare flow 9

Compare system 10

Tests

100,000 random documents processed

to vector form

Zynq software solution:

One and two cores

ARM Cortex-A9

667 MHz

Zynq PS + PL solution

8 paralel channels

100 MHz

11

Runtime comparison 12

0

5

10

15

20

25

30

35

40

45

50

55

60

0 2 4 6 8 10 12 14 16 18 20 22 24 26 28

Pro

cess

ing

tim

e [s

ec]

Number of processed reference vectors

Single core

Dual core

FPGA

Future work

Compression

High performance hardware

Higher level language

13

Text comparison 14

Cluster 15

Questions 16