The Bison cluster must include modern 64-bit x86 CPU technologies, 80 (or more) Nvidia H200 141GB GPUs, and 20 (or more) Nvidia L40S 48GB GPUs, or better. The H200 compute nodes are intended for highly arithmetic-intensive computation and for AI training/fine-tuning, and the L40S compute nodes intended for interactive AI inference. CPUs must report base clock speeds of 2.0 GHz or greater. All nodes must support all high speed and administrative networks included in Fig. 1. The compute cluster must possess the following attributes: At least 128 CPU cores per node. At least 4GB/core of memory of 4800MHz R-DIMM performance or better. Ideally, all memory channels should be populated. At least 4 PCIe-connected Nvidia H200 GPUs per chassis 10G Base-T (or better) network adapter (at least 2-ports) for administrative networks At least one Nvidia NDR/HDR two-port card on all nodes The (6) admin nodes must possess the following attributes: At least 32 CPU cores per node. At least 4GB/core of memory of 4800MHz R-DIMM performance or better. The two login nodes and two Open OnDemand nodes should have at least 480TB of SSD storage; the two management nodes should have at least 2TB of SSD storage. 10G Base-T (or better) network adapter (at least 2-ports) for administrative networks. Nvidia NDR/HDR two-port cards on all nodes.
