Runtime comparisons for 2D models in InfoWorks ICM

We’ve recently conducted some comparative tests on a selection of different 2D models, each run on a range of different PC hardware.  The aim of the tests was to see what influence the hardware had on runtimes, and particularly the potential speed gains of using GPU technology.

2D flood spread model created in InfoWorks ICM

2D flood spread model created in InfoWorks ICM

To start with, each run was done using the PC’s main CPU processor (Intel Corer2Quad, Core i3, Core i5, Core i7, Twin-Xeon etc).  Each run was then repeated utilising the GPU card in the machine.  The tests were conducted using a low end GPU card (GeForce GT440), a mid-range GPU card (GeForce GTx580) and a high-end GPU card (TESLA 2050).  At the time of writing (December 2012) a GT440 GPU card costs approximately UK£45 (US$70, AUS$80), a GTx580 costs about UK£300 (US$470, AUS$480) and a TESLA C2050 costs around UK£1,500 (US$2,300, AUS$2400). The tests were just comparing model runtimes on the different CPU hardware and then looking at 2D runtime improvements when the calculations were pushed to a GPU.

All the models were pure 2D networks; they didn’t have any 1D components (i.e. no pipes or manholes, just a 2D mesh).  Each model was run for exactly the same simulation time (2hrs).  Rainfall was applied to 100% of the 2D mesh, so all elements in the 2D zone were wet throughout the simulation, thus ensuring the 2D calculations were being applied to every single element in the 2D mesh at every timestep of the whole simulation.

Outcomes
The results of the tests are plain to see.  Even the low-end GPU cards had a significant effect on reducing the overall runtimes.  Investing in a mid-range or high-end GPU provides dramatic reductions in run-time.

If a GPU card is not available, then using either a Core i7 or Twin-Xeon processor is important to keep 2D runtimes manageable on large meshes.

Hardware Specifications
The full specification for each PC and its associated GPU card is shown below.

Intel Core2Quad – CPU Q9550 (4 cores) @ 2.85GHz, Windows 7 x64, 6.0Gb RAM.
         GPU – GeForce GTX 580 1.5Gb GDDR5.

Intel Core i3 – CPU 550 (2 cores, 4 threads) @ 3.2GHz, Windows 7 x64, 4.0Gb RAM.
         GPU – ASUS GeForce GT 440 1Gb GDDR3.

Intel Core i5 – CPU 750 (4 cores) @ 2.67GHz, Windows 7 x64, 6.0Gb RAM.
         GPU – ZOTAC GeForce GT 440 2Gb GDDR3.

Intel Core i7 – CPU 920 (4 cores, 8 threads) @ 2.67GHz, Windows Vista x64, 8.0Gb RAM.
         GPU – TESLA C2050 3Gb GDDR5 GPU.

Twin Intel Xeon – CPU E5645 (12 cores in total) @ 2.40GHz, Windows 7 x64, 24Gb RAM.
         GPU – TESLA C2050 3Gb GDDR5 GPU.

Finally, it’s very important to note that a GPU card plays no part at all in regular 1D calculations.  To get the best performance on a 1D model you simply need to use a PC with a fastest processor you can find/afford.  You may like to use this website to compare performance of the various processors that are currently available -> http://www.cpubenchmark.net/cpu_list.php.  It’s not a definitive guide, but they do have benchmarking data for most of the common AMD and Intel processors, which can be compared with those going back two or three years.

 

Share this post!

    About Andrew Walker

    Andrew Walker is a Senior Client Services Manager with Innovyze in the United Kingdom, specializing in the computerized analysis of drainage and flooding. He has over 30 years experience of modeling the key hydraulic processes involved in urban drainage design and analysis. He is one of the key members of staff tasked with supervising the roll-out and adoption of InfoWorks ICM throughout the UK and Europe.
    This entry was posted in InfoWorks ICM and tagged , , . Bookmark the permalink.

    One Response to Runtime comparisons for 2D models in InfoWorks ICM

    1. An exciting post, Andrew! Looking the times, it looks like increasing from 2 to 12 cores speed up computations by a factor of 5 and having a higher end GPU card is twice to ten times faster? It also looks like the larger the model, the faster the speedup.

    Comments are closed.