Germany – Computer platforms – Beschaffung von (direkt-wassergekühlten) HPC- und KI-Cluster-Komponenten

A directly hot-water-cooled system complex consisting of necessary racks, management servers, network components, and accelerated compute nodes with two CPUs using an x86_64 compatible instruction set and four high-end GPUs per node. The compute nodes will be networked via InfiniBand. | A directly hot-water-cooled system complex consisting of necessary racks, management servers, network components, and general-purpose compute nodes based on an x86_64 compatible instruction set. The compute nodes will be networked via InfiniBand. | A directly hot-water-cooled system complex consisting of necessary racks, management servers, network components, and AI compute nodes with at least four high-end GPUs per node. The compute nodes will be networked via InfiniBand. A direct hot-water-cooled system complex consisting of necessary racks, management servers, network components, and infrastructure to enable operation (hot-water cooling infrastructure, power distribution, etc.), as well as AI compute nodes with at least four high-end GPUs per node. The compute nodes are to be networked via InfiniBand. In contrast to lots 1 to 3, the NUTS code applicable to the place of performance is DE252. | A turnkey, AI-optimized HPC standalone solution with approximately 1,200 GPUs and 10 PB NVMe storage, housed in containers for installation of the entire system on an open area, and infrastructure to enable operation (hot-water cooling infrastructure including cooling tower, power distribution, etc.). The accelerated compute nodes are to contain at least four high-end GPUs and be connected via an InfiniBand network. External connectivity is provided via 100 GbE trunks. This is not a framework agreement, but a system supply contract. Contrary to Lots 1 to 3, the relevant NUTS code for the place of performance is DE252. | Not necessarily directly warm water-cooled, throughput-optimized GPGPU nodes with at least 4 mid-range GPUs (RTX PRO 6000 Blackwell or equivalent). The compute nodes are to...