Milestone: Training POC for 80B-Parameter Brain Model on Marlowe Completed
Professor Dan Yamins’ team completed a proof-of-concept hero run on Marlowe, validating that an 80-billion-parameter brain-inspired neural network can train at scale across 24 nodes (192 H100 GPUs). The 24-hour run achieved 45% model FLOPS utilization (MFU), confirming that capability-scale computational neuroscience training is feasible on Stanford’s GPU computational instrument.
The Counterfactual World Model is designed to study how biological neural circuits give rise to cognitive function. A dedicated 30-day training campaign for PSI2-30B, a 30-billion-parameter model, is planned as the first hero run on Marlowe.