Organizations across industries are racing to unlock the transformative potential of Generative AI (GenAI) and large language models (LLMs). This includes long-standing technology leaders who have pioneered AI research for decades and are now focused on building large-scale AI surrogate models to accelerate scientific discovery and scale engineering innovation.

The motivation is clear. Traditional physics-based HPC simulations often demand massive supercomputing resources — and even then, some workloads can take days, weeks, or months to complete. GenAI offers a breakthrough: the ability to learn from these models and dramatically reduce runtime, potentially enabling complex simulations to run on modest hardware — even laptops.

This capability is especially critical for a class of HPC workloads known as urgent computing; time-sensitive applications that require immediate access to computational resources for real-time decision-making during crises such as natural disasters, public health emergencies, or geopolitical disruptions. Examples include storm surge forecasting, earthquake early warning systems, and pandemic modeling for drug discovery and vaccine development.

At the heart of this innovation is Thunderbird, a “supercomputer-cluster-on-a-chip” designed to support the development of high-fidelity AI surrogate models. In science and engineering, precision matters — these problems demand the right answer, not just a faster approximation. That’s why every iteration of an AI surrogate model must be validated against high-precision, physics-based HPC simulations running in parallel. This ensures the surrogate model remains grounded in reality and avoids the risk of “hallucinated” outputs.

Thunderbird’s architecture enables this hybrid workflow — combining the speed and adaptability of AI with the rigor and reliability of classical simulation — to deliver trustworthy acceleration for the most demanding scientific and engineering challenges.

• Versatile all-CPU architecture applicable to all HPC-AI software, much of which does not benefit from GPUs

• High precision, high performance, low power 64-bit native processors to solve “big math” problems required by most HPC software