GPU processing with OptiSystem [Optical System]

[Damian Marek]

Question:

I have been enabling CUDA support in my OptiSystem projects, but I would like a little more information on how enabling this function speeds up simulations?

[Damian Marek]

Answer:

The GPU processing of OptiSystem is compatible with the CUDA architecture that has been developed by NVIDIA (http://www.nvidia.ca/object/cuda_home_new.html) . The technique involves using the graphics processing card (GPU) of the computer to run multi-threaded calculations through a series of parallel cores in the GPU.

It is beneficial if an OptiSystem design has long fiber links or the design requires the frequent transition from the time domain to the frequency domain (FFT) and back (IFFT) (for example if you have lots of filters in the design). This function can be handed over to the GPU instead of being processed by the computer’s CPU. For fibers it is useful as the split step Fourier technique requires the frequent use of the FFT operation.

The increase in performance is highly variable as it depends on the type of design and the number of GPU cores available on the host computer.

The “Graphics Processing Unit (GPU) computing” section on pages 5-7 of the OptiSystem User Reference guide provides details on how to setup the feature (please note that the host computer must have an NVIDA card installed and also must have the ability to support CUDA – which is generally the case for most recent computers).

[Dr. MohammadMahdi Ariannejad]

excuse me totally if I want to use bpm and optiFDTD which specification would be a good computer? I mean graphic card and CPU and ram as well as you suggest me to use PC or Laptop? many thanks. I am using core i3 dell and unfortunately all of the times I am suffering from a lots of crashes and slow speed also my ram is 6 GB but the graphic card is on board Intel, I am pretty sure it is not a good graphic card, do you think because of this my computer has a lot crashes?

[Damian Marek]

Here is a post on what general type of computers to go for:

Computer Configuration [FDTD]

As for the cause of your crashing it could be memory related. Are you using OptiFDTD 32 bit or OptiFDTD 64 bit? And what version (e.g. 12.0)

Regards

[Alessandro Festa]

Hi Damian, it would be really helpful to know the estimated advantage (in terms of % of less simulation time) for a certain GPU when using CUDA.
Do you think it would be possible for Optiwave to share this kind of data? It would help the user to choose the proper workstation based on performances and prices.

[Ravil]

It would be a great design advantage to estimate the processing time for given CPU and graphic memory of the workstation.

[Alessandro Festa]

Yes Ravil it would be good to know if spending 3k dollars is giving you a real advantage…Damian do you have any test made at Optiwave with different graphic cards?

[Damian Marek]

Hi Alessandro and Ravil,

These are good points, unfortunately I don’t think we have any rigorous tests comparing different graphics cards. I’ll see what I can find and if we don’t have any I might be able to do the tests myself.

However, the CUDA processing primarily speeds up Fourier transform calculations, so the increase in speed will depend on the project files themselves. For example, projects with large optical fiber links that include dispersion and nonlinearity should be sped up considerably. It also might be hard to benchmark all graphics cards with our resources, but I could definitely do it with the cards we have available here.

Regards

[Damian Marek]

Update on the GPU benchmarking:

I forgot to mention that we do have samples for testing the speed of CUDA processing at OptiSystem 13 Samples -> OptiSystem performance -> CUDA examples. Although without a graphics card that specializes in scientific computing, you will probably experience 5 to 6 times increase in performance.

Currently I am arranging for a test with the NVIDIA Tesla K40 and if all goes well we should have a good benchmark for what is possible with CUDA.

Cheers

[Heitor Galvao]

The advantage is very good of CUDA, for more simulation time optimization you need a good video card.

[Alessandro Festa]

Hi Heitor, can I ask you which card do you have?

[Heitor Galvao]

De acordo com a escala que foi proposto no tópico https://optiwave.com/forums/topic/earn-a-1000-for-forum-participation/ é $20

[Marcionilo José da Silva]

Hi,

I wonder if the simulation speed increased considerably if an NVIDIA GPU 384-bit 12GB with 3072 CUDA core is used instead of a 256-bit 4GB with 1664 CUDA core.
What is the relationship between memory and number of cores and the simulation time?
It is worth investing in a high performance GPU?
What is the most appropriate NVIDIA for Optisystem 13.0 (the best possible)?

Currently, we simulate a system with transmission of 40 channels at a link with 300km with dispersion and nonlinearity.

[Author]

Posts