Sampling- and Estimation-based Strategies for Data Collection in Wafer-Scale Evolution Simulations
| Authors | Matthew Andres Moreno |
| Date | March 5th, 2026 |
| Venue | Neocortex Seminar Series, Pittsburgh Supercomputing Center |
Abstract
Emerging AI/ML-oriented hardware accelerators, like the 880,000-processor Cerebras Wafer-Scale Engine (WSE), have potential to open new frontiers in computational modeling through orders-of-magnitude scale-up of high-performance computing (HPC) workloads. In the context of evolutionary biology, these technologies offer new opportunities for digital experiments exploring cross-scale biological phenomena — such as many-species eco-evolutionary dynamics and evolutionary transitions in individuality (e.g., multicellularity, eusociality). Effectively harnessing AI/ML accelerators for scientific computing workloads, however, poses substantial engineering challenges. One such challenge is tracking simulation dynamics that take place across a vast, highly-distributed fabric of memory-constrained processors. This talk will present technical and practical aspects of sampling- and estimation-based data collection strategies developed to support digital evolution on the Wafer-Scale Engine. At scale, these strategies enable the tracking of evolutionary history across trillions of simulated organisms in agent-based models. The talk will also review recent work migrating experiment and data management pipelines for general-purpose, SDK-based Wafer-Scale computing to the Cerebras Wafer-Scale Cloud.