Summary
Lenovo has outlined research-computing deployments spanning climate modelling, fusion simulations, astronomy, drug discovery and MRI research. The company reports substantial performance gains in several cases, with results tied to particular workloads.
Lenovo has described how research organisations use its high-performance computing (HPC) systems and AI infrastructure for work ranging from climate modelling to fusion research. In a company-authored overview dated 9 November 2026, Lenovo reports performance gains at several institutions; the figures refer to different systems, workloads and comparisons rather than one common test.
HPC systems combine processors, accelerators, memory and storage to handle demanding calculations in parallel. Research centres increasingly use the same infrastructure for simulations and data analysis as well as AI training—the process of fitting a model to data—and inference, when a trained model is used to generate results.
Reported gains across research workloads
At Germany’s Potsdam Institute for Climate Impact Research, Lenovo says a newer supercomputer provides five times the computational capacity of its predecessor. Depending on the workload, the institute saw a reported performance improvement of 5–10 times. The work includes repeated numerical simulations with changing parameters, which can help researchers examine climate scenarios and develop policy guidance.
For the Korea Meteorological Administration, Lenovo describes a system with more than 300,000 Intel Xeon Scalable processor cores and a stated performance of 50 petaflops. A petaflop is one quadrillion floating-point operations per second. Lenovo says the system supports decade-scale climate modelling and weather forecasts at higher resolution than was previously possible.
At ENEA, Italy’s national research body, the CRESCO8 cluster raised the capacity of the CRESCO platform from 2 to 9 petaflops. Lenovo reports that fusion-simulation workloads run five times faster than before, supporting research to refine reactor designs. In Chile, a Lenovo system used by researchers analysing telescope data is reported to have doubled research performance per watt, helping them process astronomical events more quickly.
Other examples concern AI workloads. Lenovo says a University of Queensland research group reduced AI model-training time from four days to 5.5 hours. The group is investigating ways to speed up MRI scans and make them more affordable; the reported training-time reduction is an infrastructure result, not a clinical outcome. At the University of Florida, Lenovo reports that the HiPerGator 4th Gen system provides 17.49 petaflops of compute performance, with three times faster AI training and 15 times faster AI inference than the comparison system.
Cooling and access to compute
As processors and accelerators become more densely packed, the heat they generate becomes a system-design constraint. Lenovo presents liquid cooling as one way to remove heat directly from hardware. The company says its sixth-generation Neptune liquid-cooling technology can support racks above 100 kilowatts and reduce power consumption by up to 40%; these are Lenovo-stated capabilities. At Berlin’s Zuse Institute, Lenovo says direct heat removal enables support for next-generation AI workloads. The institute’s research includes training neural-network force fields for simulations of large biomolecules, with the longer-term aim of improving understanding of protein behaviour.
The overview also describes a GPU cloud built with the Mass Open Cloud Alliance. Its pay-as-you-go model is intended to let researchers access GPU computing without buying their own hardware. That approach can make shared infrastructure available to more research groups, while liquid cooling helps manage the heat from dense server configurations.
The examples show how HPC infrastructure can serve both simulation-heavy research and newer AI workloads. Lenovo’s reported gains are specific to the named institutions and comparisons: the source does not give a shared test protocol across the projects, so the figures should not be read as directly comparable measures of one system’s advantage over another.