Case Study: Lawrence Livermore National Laboratory achieves 18 million DNN inferences per second with Cerebras CS-1

Lawrence Livermore National Laboratory - Customer Case Study

Lawrence Livermore National Laboratory (LLNL), a U.S. Department of Energy research facility focused on national security and nuclear fusion science, needed a way to speed up expensive inertial confinement fusion simulation experiments. Its HYDRA multiphysics code uses a CRETIN module that can consume a significant share of total compute time, so LLNL sought a faster way to run deep neural network inference inside the simulation loop. Cerebras provided the CS-1 system to support this high-throughput inference workload.

Cerebras integrated the CS-1 with LLNL’s Lassen supercomputer over InfiniBand and worked with LLNL to connect HYDRA to the CRETIN-surrogate model via a C++ API. The result was 1.2 Tbps bandwidth to the accelerator and inference throughput of 18 million samples per second—37x faster than a single Lassen GPU and 16x faster than a full four-GPU compute node. LLNL also reported a 5x performance improvement per transistor versus GPUs, enabling previously intractable experiments at lower cost and with much simpler integration.

Lawrence Livermore National Laboratory

Brian Spears

Principal Investigator