Researchers across the world will soon have access to a new leading-edge AI compute technology with the installation of Graphcore’s latest Bow Pod Intelligence Processing Unit (IPU) system at the U.S. Department of Energy’s Argonne National Laboratory.
The 22 petaFLOPS Bow Pod64 will be made available to the research community via the Argonne Leadership Computing Facility’s (ALCF) AI Testbed, which provides access to some of the world’s most advanced AI systems for scientific research.
With the ALCF AI Testbed, Argonne is leading efforts to integrate traditional high performance computing (HPC) resources with emerging AI accelerator technologies to support data-intensive research and next-generation scientific machine learning (ML) workloads.
The testbed systems are enabling pioneering research at the intersection of AI, HPC, and big data, including studies involving climate predictions, drug discovery, and the analysis of large-scale experimental datasets.
“Novel processor architectures like the IPU are facilitating and accelerating new AI techniques and model types. We are excited to work with the research community on ALCF’s latest Bow Pod system to advance AI for science,” said Venkatram Vishwanath, Data Science Team Lead at the ALCF.
The addition of Argonne's second IPU Pod follows an earlier, successful evaluation of Graphcore technology which concluded that the IPU is “well suited for common ML workloads and irregular workloads.”
As part of the study, the team ran scientific ML applications, BraggNN (data analysis for X-ray experiments) and CANDLE Uno (cancer research), on the Graphcore IPU-M2000 system, demonstrating its capabilities to enable highly effective distributed implementations.
The team will present a paper, “A Comprehensive Evaluation of Novel AI Accelera with tors for Deep Learning Workloads,” detailing their findings at SC22.
Graphcore's co-founder and CEO Nigel Toon has welcomed the expansion of Argonne's IPU offering to researchers: "We designed the IPU as a tool to help accelerate innovation and our relationship with Argonne National Laboratory is helping to put Graphcore technology into the hands of some of the world's leading researchers."