Exploring the New GPU Advancements (Lumerical FDTD)
Hi, this is Majid from Ozen Engineering, Inc. Today, I would like to compare the speed of GPU and CPU for a simulation. The structure we are examining is a grading coupler. As you can see, we have a fiber and a grading structure. For more details about this structure, please refer to our knowledge base article.
Preparing the Simulation for GPU
- Select the resource. We have different options in the FDTD: GPU, CPU, and GPU. I will select GPU 0. If your computer has more than one GPU, you can add the desired number of GPUs.
- Click on the Run Test. A green button indicates that the status is okay and the simulation is ready to run.
- Note the number of streaming microprocessors: 142.
- For the FDTD part, select Edit Object and go to Advanced Options. Check the box for Express Mode.
- Set the boundary condition to PML and the mesh accuracy to 2.
- The grid points in X, Y, and Z total around 18 million.
Running the Simulation with GPU
To monitor the GPU usage, open the Task Manager and select Performance and then GPU. My GPU is an NVIDIA RTX 6000. During the simulation, around 10% of the GPU is used, with dedicated GPU memory at 5 GB.
Once the simulation is finished, a Notepad file is generated containing the simulation results. The total simulation time is 448 seconds, and the FDTD solver speed for one microprocessor is 3671 million notes per second.
Running the Simulation with CPU
My CPU is a Core i9 with 16 cores. To configure the simulation:
- Select the CPU and change the core number to 16.
- Click on Run Test.
- For the FDTD part, use the same mesh accuracy of 2 and boundary condition. Disable Express Mode as it is used for GPU.
During the CPU simulation, CPU usage is at 100%. The maximum time remaining is 4 hours.
Comparison of Results
The comparison between GPU and CPU simulations is as follows:
- GPU Total Time: 448 seconds
- CPU Total Time: 2736 seconds
- The GPU is approximately 6 times faster than the CPU.
- FDTD Solver Speed:
- GPU: 3671 million notes per second
- CPU (1 core): 22.53 million notes per second
- CPU (16 cores): 360 million notes per second
In conclusion, the GPU solver is significantly faster than the CPU. Thank you for watching.
Hi, I'm Majid from Ozen Engineering. Today, I'd like to compare the speed of a GPU and a CPU for one simulation. The structure I want to compare is a grading coupler, as shown here. We have a fiber, a grading structure, and another fiber.
For more details about this structure, please refer to our knowledge base article. To prepare the simulation for the GPU, first, select the resource. We have a GPU and a CPU option in FDTD. I will select the GPU (0), but if you have more than one GPU, you can add the number of GPUs you wish.
Next, select the "Run Test" button. If the status is okay, you will see a green button. The number of streaming microprocessors for this GPU is 142. For the FDTD part, select FDTD, then "Edit Object." In the advanced options, check the "Express Mode" box.
The boundary condition is PML, and the mesh setting accuracy is two. The number of grids in x, y, and z is around 18 million grid points. Now, run the simulation with the GPU. You can check the GPU usage in the task manager.
I am using an NVIDIA RTX 6000. The GPU memory dedicated to this simulation is 5%. The percentage of CPU used in this simulation is 10%. The simulation is now finished. You can look at the notepad output. The GPU information includes relation size, grid points, and simulation size.
The total simulation is 448, and the FDTD solver speed for one microprocessor is 3671 million notes per second. Now, let's run the simulation with the CPU. I have 16 cores (Core i9). Configure the resource side by selecting the CPU and changing the number to 16. Run the test again.
For the FDTD side, we have a mesh accuracy of two, the same boundary condition, and have disabled express mode. Run the simulation with the CPU. The CPU usage is now 100%. The maximum time remaining is 4 hours. Comparing the results, the GPU is almost 6 times faster than the CPU.
The FDTD solver speed for one process is around 3671 million notes per second, while for 16 cores, it's about 360 million notes per second. Thank you.
