New Supercomputer at KIT

With the delivery of the first racks a few weeks ago, work has begun on KIT's new supercomputer, HoreKa 2.

The system will be the successor to the HoreKa high-performance computer and will provide researchers with additional computing capacity. The aim is to meet the ever-growing demands for computing power and data processing in science.

The new infrastructure relies on modern, highly efficient components to ensure maximum performance with optimal energy efficiency.

Lenovo SD665 V3 systems with the latest AMD Turin technology are used in the CPU partition. All components are supplied by a complete hot water cooling system, which enables maximum density and efficiency.

The GPU partition is based on the Lenovo ThinkSystem SC777 V4, which combines ARM-based NVIDIA Grace CPUs and NVIDIA Blackwell GPUs in a compute tray – connected via NVLink. This is currently the most powerful platform for GPU and AI applications. Here, too, Lenovo Neptune technology ensures complete direct hot water cooling of all components.

The storage system is based on IBM Storage Scale Systems (SSS 6000) with a two-tier architecture consisting of NVMe and HDD tiers. This NVIDIA-certified high-performance solution forms the basis for HPC and AI installations on NVIDIA Grace Blackwell systems, thereby promoting the next generation of data-driven research and innovation.

Where technically possible, the management systems are also water-cooled and built on Lenovo Neptune Core (OpenLoop) technology to ensure a consistently sustainable and efficient overall system.

Thanks to numerous water-cooled components, the new supercomputer sets new standards in terms of efficiency and energy consumption.

99% of the heat generated is dissipated directly into warm water, enabling highly efficient cooling throughout the year. The waste heat generated is used to heat the buildings on site.

The construction of HoreKa 2 is divided into two phases.

Phase I involves the installation of the storage and CPU units. User operation is planned to start in early 2026.

This year, Phase II will see the installation of the GPU units and the transition to full operation with full system performance.