CUDA Version (Compute Capability)

[HYParticipant]

CUDA comes with many versions (or compute capability). May I know up to which CUDA version that OptiSystem 13 is capitalizing on? I have limited budget on GPU and is looking for consumer grade (Geforce) GPUs. Should I put my money on GPU that supports higher CUDA version (2.1 versus 3.0) or GPU that comes with more memory (1GB versus 2GB)?

Many thanks.

[RavilParticipant]

Hi HY, what kind of CPU did you use with CUDA 2.1?

[HYParticipant]

Hi Ravil, I have yet to decide on which GPU to buy. I am currently considering the following:

Geforce 210 (CUDA 2.0)
Geforce GT610 (CUDA 2.1)
Geforce GT630 (CUDA 2.1)
Geforce GT640-GDDR3 (CUDA 2.1)
Geforce GT650 (CUDA 3.0)
Geforce GT640-GDDR5 (CUDA 3.5)

As you can see, these GPUs support different CUDA versions (means they have different compute capability). Moreover, they come with various data width (64 vs 128 bits) and memory size (1GB, 2GB and 4GB).

I would like to know which factor (CUDA version, data width or GPU memory size) is the most vital in speeding up the simulation of IFFT/FFT operation in Optisystem.

[Damian MarekParticipant]

We support the most up to date version of CUDA, so any modern NVIDIA card will do. Cards engineered for scientific calculation will be fastest (double arithmetic) like the NVIDIA TESLA.

[RavilParticipant]

Thanks for your elaboration, Damian! Do you think that, in general, the memory size will be more important or the the width of data bus?

[Damian MarekParticipant]

We haven’t done a verification of this and I am not a computer engineer, but I would say they are equally important. If you cannot hold the entire signal in the card then it will slow down, and if you cannot carry this signal in one chunk to the memory than it will also be slow.

[HYParticipant]

Thanks to Damian and Ravil, I will try to get funding for a Tesla-based workstation in office. But for home, I can only settle for a mid-range Geforce.

It will be a great help if anyone can rank the importance of the following for Optisystem simulation :
– CUDA Version
– Memory Size
– Data bus Width
– Memory type (DDR3 vs DDR5)

[RavilParticipant]

Thanks, Damian! I see your point.

[Alessandro FestaParticipant]

I am also interested in this topic. I currently have a GTX460, which is CUDA enabled but does not give great improvement in speed for OptiSystem simulations

[HYParticipant]

According to info from NVidia website https://developer.nvidia.com/cuda-gpus, GTX460 supports CUDA version 2.1.

Referring to http://en.wikipedia.org/wiki/CUDA, it seems to suggest that CUDA 2.1 only uses 48 cores for integer and floating-point arithmetic functions operations. (Correct me if I am wrong)

I wonder whether the dismay result of GTX460 is due to the low CUDA version, limited graphic memory or the nature of the simulation performed.

[Fairuz AbdParticipant]

Hi all,

I have GT-640 and Optisystem 13. However, CUDA enable run cause the program to crash and close. Dumbfounded.

[Damian MarekParticipant]

Can you try uploading your project file, so I can try it?

Regards

[Fairuz AbdParticipant]

I run the sample provided in CUDA folder.