Skip to main content

Accelerating FDTD Simulations with GPU Power in Lumerical's 2023 R2 Release

In this blog, we will explore the integration of GPU acceleration in Lumerical's FDTD (Finite-Difference Time-Domain) simulations, made by the Lumerical 2023 R2 release. Discover how utilizing the immense computational capabilities of NVIDIA GPUs can revolutionize your engineering and design workflows.

Releasing Parallel Processing Power

At the core of GPU acceleration lies the parallel processing prowess of modern GPUs. Lumerical's FDTD simulations are inherently complex, but GPUs excel at handling multiple tasks simultaneously. This parallelism is the key to significantly reducing simulation times, a stark contrast to the sequential nature of CPU-based processing.

Hardware Requirements: NVIDIA GPUs with Lumerical

To tap into the full potential of GPU acceleration, you'll need an NVIDIA GPU equipped with CUDA support, seamlessly integrated with Lumerical's simulation environment. CUDA enables Lumerical's software to leverage the GPU's computational might.

Ensure your NVIDIA GPU meets specific hardware requirements, including a Compute Capability greater than or equal to 3.0, and confirm that Unified Memory is accessible and that you have the correct driver versions installed. For more information we recommend referring to Ansys Lumerical GPU guide( FDTD GPU Solver Information)

Navigating Licensing Considerations

GPU acceleration introduces licensing considerations similar to those for CPU-based simulations. Streaming Multiprocessors (SMs) on the GPU are treated as CPU cores concerning licensing. Therefore, it's crucial to ensure that you possess the necessary licenses to match the number of SMs in use.While running multiple simulations concurrently may demand multiple licenses, our findings suggest that running simulations sequentially can be just as time-efficient.

Real-World Performance Gains

Let's delve into the tangible performance enhancements achievable through GPU acceleration with Lumerical's FDTD simulations. The table below showcases these remarkable improvements:

 

Mesh Size (Cells)

Hardware

Time (s)

FDTD Speed on Process (Mnodes/s)

FDTD Speed (Mnodes/s)

Improvement

(1757x837x62)

Nvidia RTX 6000

1712.06

5757.2

5757.2

 
 

Intel Core i9-7960X (16 cores)

12841.6

23.8356

381.369

750.07%

           

(1368x651x51)

Nvidia RTX 6000

936.299

4648.52

4648.52

 
 

Intel Core i9-7960X (16 cores)

5898.37

23.1398

370.236

629.97%

           

(977x466x39)

Nvidia RTX 6000

448.447

3671.24

3671.24

 
 

Intel Core i9-7960X (16 cores)

2736.51

22.5369

360.591

610.22%

 

This table serves as a testament to the power of GPU acceleration, showcasing substantial performance improvements across various simulations within Lumerical's FDTD.

Navigating Current Limitations and Anticipating Future Potential

While GPU acceleration in Lumerical's FDTD simulations offers significant advantages, it's essential to acknowledge its current limitations. Certain boundary conditions, sources, and monitor types may not yet be fully supported. For a comprehensive understanding of these limitations, we recommend referring to Lumerical's official documentation on the FDTD GPU Solver Information.

Visual Insight: For a direct and dynamic understanding of GPU acceleration's impact, we've prepared a video demonstration. This visual resource vividly illustrates the efficiency gains achieved with GPUs in FDTD simulations. Watch the video to see GPU technology in action and explore its potential for accelerating your own engineering and design tasks.

 

Conclusion

In this exploration of GPU acceleration in FDTD simulations, we've witnessed a remarkable shift in computational efficiency. From the technical details to real-world comparisons, it's clear that GPUs are rewriting the rules of simulation speed. With this newfound power, engineers and designers can approach their tasks with newfound agility, ushering in a new era of innovation.

Post by Majid Ebnali Heidari
October 12, 2023