GPU Performance Tests of the

RiverFlow2D Model

January 2019
 RiverFlow2D GPU Tests

RiverFlow2D© model and documentation produced by Hydronia, LLC. Pembroke Pines, FL. USA.

Information in this document is subject to change without notice and does not represent a commitment on part of Hydronia, LLC.

RiverFlow2D. RiverFlow2D Plus and RiverFlow2D GPU are copyrighted by Hydronia, LLC. 2011-2019.
All other products or service names mentioned herein are trademarks of their respective owners.

All rights reserved. No part of this publication may be reproduced, stored in a retrieval system or transmitted
in any form or by any means electronic, mechanical, photocopying, recording or otherwise, without the prior written permission
of Hydronia, LLC.

Technical documentation prepared by Asier Lacasta and Reinaldo Garcia.

Last document modification date: January, 2019.

Technical Support: [email protected]

Web site: www.hydronia.com

ii
RiverFlow2D GPU Tests

Contents
CONTENTS.................................................................................................................................................. III
LIST OF FIGURES ......................................................................................................................................IV
LIST OF TABLES .........................................................................................................................................V
1 INTRODUCTION .................................................................................................................................. 1
2 TEST CASES ....................................................................................................................................... 2
2.1 TEST 1................................................................................................................................................................................ 2
Test 1 Results ............................................................................................................................................................... 3
2.2 TEST 2................................................................................................................................................................................ 4
Test 2 Results ............................................................................................................................................................... 5
2.3 TEST 3................................................................................................................................................................................ 6
Test 3 Results ............................................................................................................................................................... 7
2.4 TEST 4................................................................................................................................................................................ 8
Test 4 Results ............................................................................................................................................................... 9
3 COMMENTS....................................................................................................................................... 10

iii
 RiverFlow2D GPU Tests

List of Figures
Figure 1 RiverFlow2D Plus triangular-cell mesh. .............................................................................................................................. 1
Figure 2 Main window at the end of the simulation for test 1 mesh 3. .............................................................................................. 2
Figure 3 Test 1: Speed up of the GPU solution compared against the non-parallelized CPU version. ............................................ 3
Figure 4 Main window at the end of the simulation for test 2 mesh 3. .............................................................................................. 4
Figure 5 Test 2: Speed up of the GPU solution compared against the non-parallelized CPU version. ............................................ 5
Figure 6 Main window at the end of the simulation for test 3 using Tesla K40 (top) and using Tesla K80 (bottom). ....................... 6
Figure 7 Test 3: Speed up of the GPU solution compared against the non-parallelized CPU version. ............................................ 7
Figure 8 Main window at the end of the simulation for test 4 using Tesla K40 (top) and using Tesla K80 (bottom). ...................... 8
Figure 9 Test 4: Computational cost (in seconds) for the Tesla K80, Tesla P100, Tesla V100 and RTX 2080 Ti devices. ............ 9

iv
RiverFlow2D GPU Tests

List of Tables
Table 1 Technical specification summary of NVIDIA GPU hardware. .............................................................................................. 1
Table 2 Test 1: Run times for RiverFlow2D in different GPU hardware. Intel CPU corresponds to the non-parallelized model. ..... 3
Table 3 Test 2: Run times for RiverFlow2D in different GPU hardware. Intel CPU corresponds to the non-parallelized model. .... 5
Table 4 Test 3: Run times for RiverFlow2D in different GPU hardware. Intel CPU corresponds to the non-parallelized model. ..... 7
Table 5 Test 4: Run times for RiverFlow2D in different GPU hardware. Intel CPU corresponds to the non-parallelized model. ..... 9

v
RiverFlow2D GPU Tests

1 Introduction
RiverFlow2D, is suite of two-dimensional finite-volume models for rivers, floodplains and estuaries that include flow
hydrodynamics, and add-on modules for erosion and deposition simulations, mud and debris flows, and pollutant
dispersion. RiverFlow2D can route floods in rivers and simulate inundation over complex terrain at high resolution
and with remarkable stability, accuracy and speed. The use of adaptive triangular-cell meshes enables the flow field
to be resolved around key features in complex river environments. The GPU version allows performing hydrodynamic
computations up to than 680 times faster than non-parallelized models. RiverFlow2D hydraulic simulation core has
been developed in collaboration with the Computational Hydraulics Group of the University of Zaragoza in
Spain.

This document presents several tests to demonstrate the performance of the RiverFlow2D GPU model on a
variety of real project applications using several meshes with different resolutions and utilizing various NVIDIA
GPU hardware cards (see Table 1).

Table 1 Technical specification summary of NVIDIA GPU hardware.

Tesla Tesla
Tesla K80 GTX 1080 Ti Tesla V100 RTX 2080 Ti
K40 P100

CUDA cores 2,880 2 x 2,496 3,584 3,584 5,120 4,352

Memory 12 Gb 24 Gb 11 Gb 16 Gb 16 Gb 11 Gb
Note: The sequential version of the code was run on a computer with an Intel Core i7-3820 @ 3.60 GHz CPU.

In the tests described in this document we report runtimes for each application and calculate model speed
ups with respect to the non-parallelized CPU model (using one core), which is the standard procedure to
compute speedups. For instance, if the speedup is reported to be 100, it means that the model performs
100 times faster than the non-parallelized version.

Figure 1 RiverFlow2D Plus triangular-cell mesh.

1
 RiverFlow2D GPU Tests

2 Test Cases
We present different tests to illustrate the performance of the RiverFlow2D GPU model in five real
applications using various GPU cards.

2.1 Test 1
The first test case involves the model application to a short reach of the Green River (USA) using three mesh
resolutions: 19,079 cells (Mesh 3), 154,880 cells (Mesh 3), and 1,878,607 (Mesh 3).

Figure 2 Main window at the end of the simulation for test 1 mesh 3.

2
RiverFlow2D GPU Tests

Test 1 Results
Table 2 Test 1: Run times for RiverFlow2D in different GPU hardware. Intel CPU corresponds to the non-parallelized
model.

GTX 1080 RTX 2080 Max

Mesh No. Cells Intel CPU Tesla K80 Tesla P100 Tesla V100
Ti Ti Speedup

Mesh1 19,079 00:00:08:14 00:00:00:18 00:00:00:38 00:00:00:13 00:00:00:11 00:00:00:46 45x

Mesh2 154,880 00:03:23:47 00:00:02:38 00:00:02:44 00:00:01:24 00:00:00:51 00:00:03:07 238x
Mesh3 1,878,607 08:23:17:47 00:01:28:04 00:01:08:28 00:00:33:40 00:00:18:49 00:01:00:39 687x

800.00

686.51
700.00

600.00

500.00

400.00 383.70

300.00
239.75
212.99
188.67
200.00
146.68 145.56

100.00 77.39 74.55 65.39

38.00 44.91
27.44
13.00 10.74
0.00
Tesla K80 GTX 1080 Ti Tesla P100 Tesla V100 RTX 2080 Ti
Axis Title

Mesh1 Mesh2 Mesh3

Figure 3 Test 1: Speed up of the GPU solution compared against the non-parallelized CPU version.

3
 RiverFlow2D GPU Tests

2.2 Test 2
The second test is about an application of a hydraulic structure in New Orleans at high resolution. We present
results for three meshes: 21,001 Cells in Mesh 1, 539,177 cells in Mesh 2 and 1,640,606 Cells in Mesh 3. The
project was provided by Stantec.

Figure 4 Main window at the end of the simulation for test 2 mesh 3.

4
RiverFlow2D GPU Tests

Test 2 Results
Table 3 Test 2: Run times for RiverFlow2D in different GPU hardware. Intel CPU corresponds to the non-parallelized
model.

GTX 1080 RTX 2080 Max

Mesh No. Cells Intel CPU Tesla K80 Tesla P100 Tesla V100
Ti Ti Speedup

Mesh 1 21,001 00:00:37:07 00:00:01:24 00:00:02:42 00:00:00:53 00:00:00:44 00:00:03:15 51x

Mesh 2 539,177 02:22:39:24 00:00:38:12 00:00:35:30 00:00:16:36 00:00:09:49 00:00:32:08 432x
Mesh 3 1,640,606 16:05:34:31 00:03:18:32 00:02:37:37 00:01:15:36 00:00:40:55 00:02:17:59 571x

600.00
571.27

500.00

431.86

400.00

309.19
300.00
255.39

200.00
169.40
148.30
117.74 131.93
110.98 119.42

100.00

42.02 50.61
26.51
13.75 11.42
0.00
Tesla K80 GTX 1080 Ti Tesla P100 Tesla V100 RTX 2080 Ti

Mesh1 Mesh2 Mesh3

Figure 5 Test 2: Speed up of the GPU solution compared against the non-parallelized CPU version.

5
 RiverFlow2D GPU Tests

2.3 Test 3
This test case represents an event for the simulation of a river in California (USA) including 357,611 cells. The
event covers a period of 6 days and 23 hours.

Figure 6 Main window at the end of the simulation for test 3 using Tesla K40 (top) and using Tesla K80 (bottom).

6
RiverFlow2D GPU Tests

Test 3 Results
Table 4 Test 3: Run times for RiverFlow2D in different GPU hardware. Intel CPU corresponds to the non-parallelized
model.

GTX 1080 RTX 2080 Max

No. Cells Intel CPU Tesla K80 Tesla P100 Tesla V100
Ti Ti Speedup

357,611 06:00:30:01 00:01:51:47 00:01:47:10 00:00:49:49 00:00:34:38 00:01:59:56 250x

300.00

250.34
250.00

200.00
174.04

150.00

100.00
77.56 80.90
72.29

50.00

0.00
Tesla K80 GTX 1080 Ti Tesla P100 Tesla V100 RTX 2080 Ti
Figure 7 Test 3: Speed up of the GPU solution compared against the non-parallelized CPU version.

7
 RiverFlow2D GPU Tests

2.4 Test 4
This test reports on results of an ongoing collaboration with the National Oceanic and Atmospheric
Administration (NOAA) of the USA. It shows a simulation of 420-mile reach of the Red River of the North located
in Minnesota (USA). The event involves the routing of 3-month hydrographs.

Figure 8 Main window at the end of the simulation for test 4 using Tesla K40 (top) and using Tesla K80 (bottom).

8
RiverFlow2D GPU Tests

Test 4 Results
The computer times of the non-parallelized CPU model is impractical for this test. Therefore, only the
RiverFlow2D GPU model was used.

Table 5 Test 4: Run times for RiverFlow2D in different GPU hardware. Intel CPU corresponds to the non-parallelized
model.

RTX 2080
No. of cells Tesla K80 Tesla P100 Tesla V100
Ti
4,616,546 00:20:50:46 01:02:55:42 01:02:21:04 00:12:22:54
75046

44574
33972

22948

Tesla K80 Tesla P100 Tesla V100 RTX 2080 Ti

Figure 9 Test 4: Computational cost (in seconds) for the Tesla K80, Tesla P100, Tesla V100 and RTX 2080 Ti devices.

9
 RiverFlow2D GPU Tests

3 Comments
This report presents performance results of the Riverflow2D GPU model in several NVIDIA GPUs including the
latest generation RTX, Tesla P100 and V100 cards. While the Tesla V100 is still the clear winner of the tested
devices, the NVIDIA GTX 1080 Ti card is much lower in costs and its acceleration capabilities are also
remarkable. The latest benchmarks include the RTX 2080 Ti, for which the performance gain is almost
negligible compared to the GTX 1080 Ti, therefore cannot be recommended as the best low cost solution. This
was a surprise for us since we usually see about a 20-25% increase in speed between generations.

As demonstrated in the tests presented in this document, the remarkable performance of the RiverFlow2D GPU
has several implications including:

• Computer run times are reduced from days to a few hours, or from hours to minutes, and from minutes
to seconds in some cases.
• The RiverFlow2D GPU allows evaluating river flooding simulations of large river reaches that were
impractical until recently due to excessive runtimes.
• The use of GPU technology developed in the RiverFlow2D code also allows using models with large
resolution meshes involving millions of cells.
• The emergence of Pay-per-Use Cloud Services such as the Google Cloud where all of the tested cards
are available at very attractive costs, facilitates the use of the RiverFlow2D GPU model for a wide range
or applications.

10