Karlsruhe Institute of Technology (KIT) has put HoreKa 2 into operation with a planned peak performance of around 33 petaFLOPS, more than 42,000 processor cores and 300 AI accelerators. The system gives researchers across Germany a new resource to apply for, but its headline specification is not evidence of how fast it will run AI workloads or how much computing time any team will receive.
That distinction defines what HoreKa 2 adds: a large, hybrid research system and an access route through Germany’s National High-Performance Computing network. The figures describe planned capacity, while the available material does not establish a net increase in Germany’s total computing power or measured performance on AI tasks.
A peak figure is not an AI performance result
KIT held the ceremonial commissioning at its Campus North site on Sept. 25. The institute says the project cost 17 million euros and identifies climate, materials, particle physics, energy, mobility and AI as intended research areas.
It is a measure of stated computing capacity, not a report of how quickly the machine completes a particular simulation or AI task.
KIT describes HoreKa 2’s peak as planned at around 33 petaFLOPS. Data Center Dynamics reports the same peak figure, but the material reviewed gives no benchmark conditions or independent results for AI workloads. The number therefore cannot show which models or research tasks the system will run, how fast they will run, or how its performance compares with other machines on those tasks.
The reported hardware gives a clearer picture of the system’s design. Data Center Dynamics says the CPU partition uses Lenovo SD665 V3 servers with AMD EPYC 9655 Turin processors.
Data Center Dynamics also reports 250 terabytes of memory and IBM Storage Scale storage with NVMe and hard-drive tiers. These details establish the reported configuration, not the results of a particular research run. A list of processors, accelerators and storage components cannot substitute for workload measurements.
The hybrid design is relevant to KIT’s stated aim of supporting both scientific computing and AI research. But the component list alone does not show the scale of a training run, the speed of inference, or whether a particular research group’s software and workload will perform well on the system. No independent AI workload evaluation appears in the material reviewed.
KIT Scientific Computing Center board member Martin Frank has characterized HoreKa 2 as doubling computing power for simulations and AI applications. That is KIT’s assessment; the available material does not define the comparison baseline or independently verify the claimed doubling.
HoreKa 2 is the successor to KIT’s earlier HoreKa, which entered operation in 2021. Data Center Dynamics reports that the earlier system had a 17-petaFLOPS peak at its inauguration. That historical figure does not establish the older system’s capacity when HoreKa 2 came into operation, so it cannot by itself show the net change in KIT’s or Germany’s available computing capacity.
Access will determine how much capacity researchers can use
KIT says researchers from across Germany can apply to use HoreKa 2 through the National High-Performance Computing network. The network includes nine university computing centers, and its resources are allocated through an application process. HoreKa 2 is therefore intended to serve researchers beyond KIT, though access depends on applications and available resources.
The rollout has taken place in stages. NHR@KIT said the first phase of user operations was scheduled to begin Aug. 10, initially with the Onyx CPU partition, with additional components to follow. The current cluster documentation lists Jade, the partition with Nvidia B200 accelerators, as coming soon.
That status matters alongside the 300 AI accelerators KIT reports for the system as a whole. The total does not establish that every accelerator was available to researchers at the Sept. 25 ceremony. The documentation does not give a date for Jade’s availability or specify how much computing time individual teams will receive.
Free newsletter
Get the morning briefing
Start each day with the AI stories that matter and why — a short, free email from our newsroom.
For research groups, capacity is more than a peak number. Access depends on the application process, the availability of the relevant partition and the allocation each project receives. The material reviewed does not spell out allocation limits, application criteria or a timetable for all components.
The national significance also needs careful framing. HoreKa 2 is an additional resource researchers across Germany can seek to use, and it succeeds an earlier KIT system. The supplied figures do not make a like-for-like comparison between old and new capacity, nor do they establish that the new system adds 33 petaFLOPS net to Germany’s installed computing resources.
Energy claims need workload evidence
KIT says HoreKa 2 uses hot-water cooling and that the power supply at Campus North is based entirely on renewable energy. The institute also describes the machine as among the most energy-efficient systems in its performance class worldwide. No independent assessment of that comparative ranking appears in the material reviewed.
The cooling method and electricity supply are relevant design details, but they do not establish how much energy the system uses while running a specific research task. Nor do they show how energy consumption compares with other supercomputers completing equivalent work. The available material provides no workload-level energy measurements.
The case for HoreKa 2 as an infrastructure addition rests on concrete specifications and planned access: a hybrid CPU-and-GPU system, a planned peak near 33 petaFLOPS and an application route that extends beyond KIT. Those facts make it relevant to German research. They do not yet demonstrate AI speed, guarantee access to accelerator capacity or settle how much national capacity has changed.
The most useful next evidence would be measured performance on named scientific and AI workloads, including the conditions behind those results; workload-level energy figures; and confirmation of when the Jade B200 partition becomes available. Until then, HoreKa 2’s specifications show what KIT says the system is designed to provide, not what performance every researcher can expect.