Adaptation of AI-accelerated CFD Simulations to the IPU platform

Researchers demonstrate successful adaptation of AI-accelerated computational fluid dynamics (CFD) simulations to Graphcore's IPU platform, achieving up to 34% speedup through optimized data pipeline management. The study shows strong scalability from 2 to 16 IPUs, increasing throughput from 560.8 to 2805.8 samples per second, validating IPUs as viable accelerators for AI-enhanced scientific computing workloads.

This research addresses a growing intersection between specialized hardware accelerators and machine learning applications in scientific computing. By porting AI-supported CFD simulations to Graphcore's Intelligence Processing Units, the authors demonstrate that emerging processor architectures can effectively handle domain-specific computational challenges beyond traditional deep learning tasks. The 34% performance improvement from optimizing data feeding mechanisms through the popdist library reveals that bottleneck identification and targeted optimization remain critical in heterogeneous computing environments.

The scalability findings carry significant implications for enterprise computing infrastructure. While dual-IPU configurations showed diminishing returns due to communication overhead, the linear scaling from 2 to 16 units—achieving a 5x throughput increase—demonstrates that communication costs become amortized at scale. This pattern suggests IPUs could serve latency-sensitive or throughput-intensive simulation workloads in research institutions and industrial applications requiring rapid prototyping of physical systems.

For the broader AI hardware market, this work validates a competitive alternative to GPUs for specialized workloads. As scientific computing increasingly relies on surrogate models trained through machine learning, the ability to efficiently train and deploy these models on dedicated hardware becomes economically important. Organizations running OpenFOAM or similar CFD tools could benefit from hybrid architectures combining IPUs with traditional processors.

Future developments should focus on expanding compatibility across more scientific simulation frameworks and investigating why 2-IPU configurations underperform relative to single units—a critical constraint for mid-scale deployments.

• →IPU-POD16 platform enables efficient AI-accelerated CFD simulations with up to 34% speedup through optimized data pipelines
• →Scaling from 2 to 16 IPUs improves throughput 5x (560.8 to 2805.8 samples/s), demonstrating strong inter-IPU communication efficiency
• →Dual-IPU configurations show communication overhead costs that exceed single-IPU performance, requiring scaled deployments for efficiency gains
• →Custom TensorFlow integration through Poplar SDK enables relatively straightforward adaptation of existing ML simulation code to IPU hardware
• →Specialized hardware accelerators show promise for scientific computing workloads beyond traditional deep learning applications