Removing The Accuracy And Time Tradeoff In EM Simulation

For years, electromagnetic simulation forced engineers to choose between accuracy and turnaround time.

As simulation frequencies climbed beyond 60 GHz and designs became more complex, engineers could no longer avoid mesh refinement. Higher fidelity required more mesh elements in regions of high field strength. More elements produced larger sparse matrices. Larger matrices extended solve times.

If you wanted higher accuracy, you waited longer.

That constraint shaped real engineering decisions. Teams simulated only the most critical signal nets to reduce port counts. They limited iterations. In some programs, schedule pressure drove compromises in accuracy.

That compromise is beginning to break.

Rewriting how FEM problems are solved

Finite Element Method (FEM) solvers rely on solving large sparse matrix systems. Historically, developers optimized these solvers for CPU architectures. While GPUs accelerated other electromagnetic techniques such as the Finite-Difference Time-Domain (FDTD) method, accelerating sparse direct solvers for FEM proved far more challenging.

NVIDIA’s cuDSS sparse direct solver takes a different architectural path. Its algorithms are designed specifically for massively parallel GPU execution, allowing large FEM matrix systems to run efficiently on NVIDIA H100 GPUs.

The physics does not change. The mathematics does not change. The execution model does.

Integrating GPU acceleration into a production workflow

Hardware capability alone does not improve engineering productivity. Teams must integrate it into a stable, validated design flow.

Keysight’s Design Engineering Software team collaborated with NVIDIA to integrate cuDSS directly into the FEM solver within Keysight Advanced Design System (ADS). Engineers access GPU acceleration as a native simulation option inside an established EM workflow.

Engineers can now:

• Preserve high-frequency accuracy without unacceptable runtime penalties
• Evaluate more design variants within the same schedule window
• Shift high-fidelity validation earlier in the design cycle

Keysight has long anchored its EM simulation capabilities on numerical accuracy and application-focused solvers. GPU acceleration inside ADS preserves that accuracy while scaling compute performance, allowing teams to simulate complex designs without sacrificing trusted results.

Benchmarking large-scale designs

Keysight benchmarked seven large FEM designs ranging from 5.4 GB to 80 GB using an AWS p5.4xlarge instance equipped with an NVIDIA H100 GPU.

For large sparse systems where solver time dominated overall runtime, the GPU-accelerated cuDSS solver delivered measurable performance gains compared to traditional CPU execution. Across the seven designs, speed-ups ranged from 1.5x to 6x, including a 77% solve-time reduction for the 80 GB test case.

These results do more than reduce runtime. They allow engineers to retain model complexity while compressing design cycles.

To further characterize scaling behavior, Keysight varied mesh refinement on a single design to measure linear performance trends. GPU acceleration achieved up to 2.6x speed-up as mesh refinement increased. Additional performance headroom exists as integration and workflow optimizations advance within Keysight EDA.

Fig. 1: Performance scaling of Keysight ADS FEM linear solver using NVIDIA H100 GPU. As the number of unknowns increases, GPU acceleration delivers substantial reductions in linear solve time, enabling faster high-fidelity EM simulation.

What this means for design engineering teams

As system frequencies increase and integration density grows, modeling fidelity becomes non-negotiable. The cost of insufficient simulation appears later in respins, validation delays, and performance gaps.

When high accuracy no longer requires extended wait times, engineering behavior shifts. Engineers can retain full geometry detail. They can validate higher port-count systems with confidence. They can iterate instead of defer decisions.

GPU-accelerated FEM inside Keysight Advanced Design System reduces simulation bottlenecks while preserving fidelity.

Accuracy no longer needs to be the slow option.

Fig. 2: Simulation of 3D signal nets with SIPro within Keysight’s Advanced Design System 2026 Update 2.

Keysight Design Engineering Software (DES) redefines what’s possible by closing the loop between simulation and measurement, enabling engineers to act sooner, reduce risk, and conquer complexity. Whether it’s autonomous mobility, photonics, or high-speed computing, we help bring breakthrough technologies to life—right the first time. Built on decades of measurement science expertise, the DES portfolio enables engineers to design, simulate, validate, and optimize complex systems across RF and mixed-signal electronics, mechatronics, optics, multiphysics, and software domains. By unifying design and validation workflows within an open ecosystem, Keysight helps teams accelerate development while maintaining confidence in simulation accuracy across disciplines.