Case Study: SURFsara achieves 187x faster AI radiologist training and cuts neural machine translation from months to hours with Dell Technologies

SURFsara teams with Dell Technologies and Intel to train a fast, accurate AI radiologist and to accelerate neural machine translation performance

SURFsara, the Dutch national supercomputing center, needed to help research teams accelerate deep learning projects—from medical imaging to neural machine translation—that were limited by compute, memory and training time. To tackle these challenges the organization partnered with Dell Technologies (alongside Intel) and used Dell Technologies’ HPC & AI Innovation Lab resources, including the Dell EMC Zenith supercomputer and Dell EMC PowerEdge servers equipped with Intel® Xeon® Scalable processors.

Dell Technologies deployed large-scale parallel HPC infrastructure and software optimizations to train a CheXNet-based AI radiologist on full 1024×1024 X‑rays and to parallelize neural machine translation workflows. The result: improved diagnostic accuracy for 10 of 14 thoracic pathologies (several near 90%), a reduction in AI radiologist training time from a single‑node baseline of 1,653 minutes to 8.8 minutes (187× faster on 256 nodes), and NMT training times cut from ~31 days on one node to about 6 hours (121× faster with 200 nodes), with findings shared back to the HPC and AI community.

SURFsara

Valeriu Codreanu

Senior HPC Consultant