Scientific Kernels on viram and Imagine Media Processors Manikandan Narayanan1, Leonid Oliker2 Adam JaninError: Reference source not found,3, Parry HusbandsError: Reference source not found, and Xiaoye LiError


Table 4: Percentage of algorithmic peak performance of VIRAM and Imagine for N =3, M = 1,10,20 and L =1024



Download 217.16 Kb.
Page4/4
Date09.06.2018
Size217.16 Kb.
#53784
1   2   3   4

Table 4: Percentage of algorithmic peak performance of VIRAM and Imagine for N =3, M = 1,10,20 and L =1024



Figure 4: Percentage of algorithmic peak performance of Imagine with N =3 and K=1,5,10 using long streams and varying computational intensity




VIRAM (N=3)

Imagine (N=5)

Ops/Word


50

90

120

150

100

200

300

400

% Peak

82%

88%

89%

91%

86%

89%

90%

91%

Table 5: Achieving high efficiency for VIRAM and Imagine using long streams and high computational intensity



Figure 5: Performance crossover between VIRAM and Imagine for N =3 and M =10





VIRAM

Imagine

Matrix

Rows
(Nonzeros)



Performance

CRS

Segsum

Ellpack

CRS

Streams

Ellpack

LSHAPE 1008
(6958)

% of Peak

2.8%

7.4%

31%

1.1%

0.8%

1.2%

Total cycle

66823

23802

5666

40300

48190

37930

MFlop/s

44

118

496

170

142

186

LARGEDIS

10000
(177820)



% of Peak

3.2%

8.4%

32.0%

1.5%

0.6%

6.3%

Total cycle

802070

567491

641512

742310

1840380

753540


MFlop/s

91

135

511

240

97

1088

Table 6: Performance of SPMV on VIRAM and Imagine for the LSHAPE and LARGEDIS matrices using various algorithms





VIRAM

Imagine

Matrix

Performance

MITRE RT_STAP 192-by-96
complex matrix

% of Peak

34.1%

65.5%

Total Cycles

5188817

712770

MFlop/s

546

13,100

Table 7: Performance of QRD on VIRAM and Imagine for the 192-by-96 MITRE RT_STAP matrix



1 Computer Science Division, University of California, Berkeley CA 94720

2 Computational Research Division, Lawrence Berkeley National Laboratory, Berkeley CA 94720

3 International Computer Science Institute, Berkeley CA 94704

4 Stream length ranges from 4320 to 18160 depending on the optimal strip size as predicted by the software development environment.


5 As of this time we have been unable to reproduce these results. We are currently working with the Stanford team to resolve any inconsistencies.

Download 217.16 Kb.

Share with your friends:
1   2   3   4




The database is protected by copyright ©ua.originaldll.com 2024
send message

    Main page