# Overview of Cluster The Kempner institute has dedicated compute resources for its mission toward advanced research on intelligence from biological, cognitive, and computational perspectives. ## What is the Kempner Institute AI Cluster? The Kempner cluster is a cluster of data center GPUs stacked in servers and racks, networked in such a way that the entire cluster can be utilized for both serial and distributed GPU workflows. The Kempner Institute collaborates with FAS Research Computing [FASRC](https://www.rc.fas.harvard.edu/) on the infrastructure and maintenance of this cluster. ## Who is eligible to use the Kempner Institute AI Cluster? In general terms, the cluster is reserved for individuals who are either associated with a lab that is affiliated with the Kempner Institute or those who are engaged in a project that has received approval from the Kempner Institute. ## What are the specifications of the Kempner Institute AI Cluster? The following table provides an overview of the Kempner Institute AI cluster: | Specification | A100 40GB | H100 80GB | H200 141GB | RTX6000 96GB | |---------------------------|------------------------------------------------|-------------------------------------------------------|-------------------------------------------------------|-------------------------------------------------------| | **Total GPUs** | 144 | 384 | 384 | 192 | | **Servers (per rack)** | 36 | 24 | 24 | 6 | | **GPUs per Server** | 4 | 4 | 4 | 8 | | **CPU Cores per Server** | 64 | 96 | 64 | 128 | | **RAM per Server** | 1 TB | 1.5 TB | 1.5 TB | 1.5 TB | ```{figure} figures/svg/cluster_diagram.svg --- height: 500 px name: Diagram of AI Cluster --- This is a diagram of the Kempner AI cluster within the FASRC infrastructure showing the arrangement of nodes, networking, and storage systems. ``` Currently, there are 16 GPU racks in production at the Massachusetts Green High Performance Computing Center (MGHPCC): one A100 GPU rack, four H100 GPU racks, five H200 GPU racks (one is a spare with 10 nodes and 40 H200 GPUs), four RTX 6000 PRO Blackwell GPU racks, one network core rack, and one storage rack. In the following diagram, these racks are shown along with a network core designed for non-blocking communication between GPUs. ```{figure} figures/jpg/2026_kempner-cluster_cloudbg_XL.jpg --- width: 100% name: GPU Racks and Network Core --- This is a diagram of the GPU racks and network core. ``` ```{note} The GPU network is non-blocking for the H200 and RTX 6000 PRO partitions. The H100 partition is currently 2:1 oversubscribed. ``` ## Can I buy into the Kempner Institute AI Cluster? If you have a Kempner Institute affiliation, you may request to contribute funds towards a Kempner cluster purchase. If you contribute funds we will increase your priority on the cluster to reflect the proportion of the cluster you have contributed. Please see the Cluster Governance guidelines for more information: https://kempnerinstitute.harvard.edu/kempner-community/ ## How is priority set on the Kempner Institute AI Cluster? Priority is set through Fairshare scheduling. For more information, please see the Cluster Governance guidelines: https://kempnerinstitute.harvard.edu/kempner-community/ ## Can I reserve GPUs on the Kempner Institute AI Cluster? Generally we hope to have enough GPUs available on the cluster for most use cases. However, for situations in which you need a large number of GPUs that are dedicated for a period of time, or need a smaller subset for a longer period of time, you can make a reservation request. Please see the full governance guidelines for further details: https://kempnerinstitute.harvard.edu/kempner-community/ ## How can I monitor my jobs and cluster usage? The Kempner Institute provides [KempnerInsight](https://kempnerinsight.eng.kempnerinstitute.harvard.edu/), a web dashboard for monitoring the cluster from a browser. It brings together job-level and cluster-level metrics in one place, so you can see how your jobs and the cluster as a whole are performing without running any commands. At a glance, it offers: - A searchable table of jobs, with a per-job detail view that plots CPU utilization, per-GPU compute and memory utilization, and InfiniBand throughput, so you can tell whether a job is really using the GPUs it holds. - A cluster view with a live GPU status heatmap, allocation and utilization over time, node health, and a per-node table. - Usage reports that summarize GPU hours, equivalent cloud cost, and utilization for a lab or the whole cluster over a date range. For a full walkthrough with screenshots, see {doc}`KempnerInsight Cluster Monitoring App `. For command-line and GPU-counter tools, see {doc}`Performance Monitoring and Optimization <../../s5_ai_scaling_and_engineering/efficiency/performance_monitoring_and_optimization>`. ## Where is the Kempner Institute AI Cluster located physically? The cluster is located at the [Massachusetts Green High-Performance Computing Center](https://mghpcc.org) (MGHPCC) in Holyoke, MA. This is a state-of-the-art datacenter that is shared among the Boston-area universities. It runs off of hydropower, making it one of the “greenest” computing centers in the world. FASRC also provides co-located [storage resources](https://docs.rc.fas.harvard.edu/kb/data-storage-workflow-rdm/) at MGHPCC.