High-quality 3D structure predictions for more than 41M proteins (www.llnl.gov)

Scientists from Lawrence Livermore National Laboratory, AMD and Columbia University ran ElMerFold — a protein-folding workflow on the world’s fastest supercomputer, El Capitan — to produce high-quality 3D predictions for more than 41 million proteins. The run sustained about 2,400 structure predictions per second and peaked at 604 petaflops using 43,200 AMD Instinct MI300A APUs across 10,800 nodes, completing in hours what was previously expected to take days. Through optimizations across the software stack (Flux workload manager, a persistent inference server, improved memory handling and sequence distribution) and a 17.2× speedup over the prior OpenFold2 implementation, the team converted a trained model into a massive distillation dataset faster and more efficiently than before. Technically, the effort rewrote major parts of OpenFold to exploit El Capitan’s unified CPU–GPU APU architecture (eliminating device transfers), built an LBANNv2 PyTorch backend, and used Triton and DaCe for portable kernel and spatial optimizations. The result makes training OpenFold3 — an open-source AlphaFold3 alternative — feasible at scale, enabling larger distillation datasets and eventually simulations of billions of multimers. Beyond democratizing state-of-the-art protein prediction for drug discovery, bioresilience and basic biology, these scheduling, memory and kernel innovations demonstrate how HPC + AI can accelerate other scientific domains and national‑security workflows.