TCAD For GPUs And GPUs For TCAD

It is well known that many steps in chip development become exponentially harder as feature sizes shrink and instance counts balloon. Billions of transistors are now commonplace, and wafer-scale devices with trillions are on the horizon. Such massive chips put pressure on every electronic design automation (EDA) tool in the development flow, from front-end architectural modeling to signoff and technology computer-aided design (TCAD) at the back end. Every run of every tool takes longer, more runs are required, and parallelism is mandatory.

Some of the biggest and most complex chips being developed today are graphics processing units (GPUs), which have tremendous parallel processing capability. For example, members of the NVIDIA H100 family have as many as 80 billion transistors and tens of thousands of processing cores. GPUs play a key role in the evolution of chip development because they are such important examples of large designs. In addition, GPU hardware enables TCAD and other EDA tools for developing massive chips, including new generations of GPUs.

Why industry is adopting GPUs

As the name suggests, GPUs were originally developed to render 3D graphical images quickly enough for real-time reactions to user input. Using a game controller to look around in a 3D scene requires a great deal of computation to paint realistic results. This could not be accomplished with central processing units (CPUs) alone, leading to the introduction of dedicated GPUs with modest parallelism. The number of cores in GPUs has since grown considerably, widening the gap with CPUs more for every new generation.

For this reason, usage of GPUs has grown well beyond their original purpose. Many additional software applications are adopting GPUs all the time, driven by four key factors:

• Parallel processing power: GPUs can handle thousands of operations simultaneously, making them ideal for tasks such as machine learning (ML) and simulation in addition to graphics rendering
• Performance boost: tasks that take minutes or hours on CPUs can be completed in seconds with GPUs, especially in fields such as AI, scientific computing, and big data
• Specialized frameworks: software ecosystems such as TensorFlow, PyTorch, and CUDA are now optimized to run on GPUs, making adoption easier
• Cost efficiency: for compute-intensive workloads, GPUs offer better performance-per-dollar than CPUs, especially in cloud computing environments

How TCAD helps develop GPU chips

TCAD can be a vital step in the design and optimization of GPUs. It helps simulate the behavior of transistors, such as FinFETs and GAAFETs, used in GPU cores. This allows engineers to explore different materials (e.g., high-k dielectrics) to reduce leakage and power consumption and to optimize switching speed and thermal performance. TCAD also supports process simulation (implantation, diffusion, etching, deposition, etc.), device modeling under different stress and variability conditions, and yield prediction and defect analysis.

For advanced GPUs that use chiplets or 3D stacked architectures, TCAD enables 3D and heterogeneous integration (3DHI). TCAD simulates interconnect behavior, through-silicon vias (TSVs), and bonding layers, and helps in minimizing signal loss and thermal resistance. By catching potential physical and electrical issues early in simulation, TCAD reduces the need for costly prototyping, accelerates time to market, helps to achieve first-pass silicon success, and helps to ensure that chips can be fabricated successfully with acceptable yield.

How GPUs improve TCAD simulations

Modern GPUs are built on advanced nodes, typically 3-5mm, and contain billions of transistors. Such a huge scale puts stress on many steps in the development process, including the TCAD simulations. Fortunately, many EDA tools, including TCAD, can benefit from adopting GPUs and leveraging their power:

• Massive parallelism: tasks such as SPICE simulation, parasitic extraction, and Monte Carlo analysis involve huge amounts of parallelizable calculations, which GPUs handle efficiently
• Faster turnaround time: design verification, place-and-route, and physical verification benefit from GPU acceleration, enabling shorter chip design cycles
• AI and ML integration: modern EDA tools use machine learning for design prediction, error detection, etc., which run significantly better on GPUs
• Scalability in cloud-based flows: as EDA workflows move to the cloud, GPUs offer scalable acceleration for compute-heavy steps, reducing cost and time

As shown in the graph below, GPUs become mandatory for TCAD runs on very large designs. In addition to better speed and capacity, GPUs enable advanced technologies such as design technology co-optimization (DTCO) and system technology co-optimization (STCO).

A rapidly emerging solution

GPU support for TCAD is a leading-edge technology, but it is reality, not just a future vision. The recent NVIDIA GTC AI Conference included a presentation on real-work performance results of Synopsys QuantumATK atomistic simulation software running on GPUs. In addition, results with Synopsys Sentaurus Device advanced multidimensional device simulator are showing nearly 10X improvement over CPUs across a wide range of applications, including power DTCO and 3D logic and memory. This represents the first GPU acceleration of TCAD device simulation.

TCAD on GPUs is a very active area of development. Synopsys and NVIDIA are working together in a close partnership to share expertise and best leverage their mutual technologies. Keep posted for much more news to come.