Kempner AI Cluster
The Kempner Institute’s state-of-the-art AI cluster and its dedicated team of research scientists and research engineers support advanced research on intelligence from biological, cognitive, and computational perspectives.
The Kempner Institute AI cluster opened with an initial pilot installation in Spring 2023. In October 2023, the Kempner Institute announced that it had purchased additional Nvidia H100 GPUs to be added to its AI Cluster.
The Kempner AI cluster is now fully operational with 144 Nvidia A100 40GB GPUs and 384 Nvidia H100 80GB GPUs. It is one of the largest academic AI Clusters in the world, equivalent to over 1300 Nvidia A100 GPUs in terms of FP32 TFLOPS.
The cluster uses a group of NVIDIA Quantum InfiniBand switches to create a non-blocking InfiniBand network connection between all GPUs, with a 400 Gbps InfiniBand NDR connection between H100 GPUs and a 200 Gbps InfiniBand HDR connection between A100 GPUs. The non-blocking network connection provides high-speed connectivity between GPUs to facilitate training and fine-tuning Large Language Models (LLMs). The dedicated network core provides non-blocking network connectivity for up to seven H100 racks, totaling 672 H100 GPUs.
The computing power of the Kempner AI cluster enables training of the popular Meta Llama 3.1 8B and Meta Llama 3.1 70B with one trillion tokens in about two weeks and three months, respectively.
Green High Performance Computing
The Kempner’s AI cluster resides in the Massachusetts Green High Performance Computing Center (MGHPCC), a modern datacenter in western Massachusetts shared by the Boston-area universities.