performance evaluation and benchmarking of modern gpu ...cs.tu-sofia.bg/conf/virtual/vs_02.pdf ·...
TRANSCRIPT
Text
Performance Evaluation and
Benchmarking of Modern GPU
Architectures
Desislava Ivanova, Vladimir Kadurin, Yanko Belov
Technical University of Sofia, Computer Systems Department, Sofia, Bulgaria
GPU advantages
• The advantage of heterogeneous modern
computing systems based on GPUs for
parallel processing is undisputable
• GPGPUs have many smaller "cores"
which do less work than a conventional
CPU core
• The heavy parallel work is done by the
GPU cores
• GPUs offer great performance/consumed
power ratio
• Many of the already existing codes
optimised for serial execution can be
tweaked to take advantage of the GPU
parallel processing power
GPU Architecture
The main GPU used for benchmarking is the GTX
Titan
It is Kepler based with code name GK110
GTX Titan has 15 SMX units each with 192 CUDA
cores, for total of 2688 CUDA cores
It has 6GB of GDDR5 memory and bandwidth of
288GB/s
GPU Technologies
Dynamic parallelism
Hyper - Q
Grid Management Unit
GPUDirect
• NVIDIA GeForce GTX 680
AMD RADEON HD7970
Comparison GPUs
Benchmark programs
Computational benchmarks - GEMM and FFT
Synthetic benchmarks - 3DMark, 3DMark11, Unigine
Heaven
3D games - Battlefield 3 at 4K resolution
Experimental Framework
Specially selected
components so no bottleneck
occurs
CPU and RAM are high level
parts
All GPUs are overclocked
All benchmarks are completed
several times so all external
factors can be minimised
Experimental results -
GEMM
3189
2405
1258
1324
702
124
0 1000 2000 3000 4000
GTX Titan
HD7970
GTX 680
TFlops
SGEMM DGEMM
Experimental results - FFT
455
277
229
60
0 125 250 375 500
GTX Titan
GTX 680
GFlops/s
FFT Single precision FFT Double precision
Experimental results -
3DMark Fire Strike
8786
14002
6422
9021
7106
9888
17980
7098
10743
7801
0 4500 9000 13500 18000 22500
GTX Titan
GTX TitanSLI
GTX 680
GTX 680 SLI
HD7970
Score
Score Graphics Score
Experimental results -
3DMark Extreme
4923
9042
3399
6287
3266
4587
8811
3087
6042
2965
0 2500 5000 7500 10000
GTX Titan
GTX Titan SLI
GTX 680
GTX 680 SLI
HD7970
Score
Score Graphics Score
Experimental results -
Unigine Heaven 4.0
77
129
48
90
50
65
108
41
79
40
0 32.5 65 97.5 130 162.5
GTX Titan
GTX TitanSLI
GTX 680
GTX 680 SLI
HD7970
FPS
Normal Extreme
Experimental results -
Battlefield 3 4K resolution
25
53
15
30
17
36
61
22
39
25
0 17.5 35 52.5 70
GTX Titan
GTX Titan SLI
GTX 680
GTX 680 SLI
HD7970
FPS
Min FPS Average FPS
Experimental results -
power consumption
72
87
79
96
86
371
554
312
487
377
0 150 300 450 600
GTX Titan
GTX Titan SLI
GTX 680
GTX 680 SLI
HD7970
WATTS
Standby Heavy load
Conclusion
Kepler architecture used in NVIDIA GeForce Titan
offers great performance and efficiency
Its advantages are significant
Despite providing such computing power, it consumes
a lot less energy than previous architectures
Thank you for your
attention!
FOR MORE INFORMATION, PLEASE CONTACT:
Assist. Prof. PhD Desislava Ivanova