Bull secures the next-generation Lumi AI supercomputer contract ahead of HPE.

Europe’s approach to generative AI infrastructure remains notably more traditional than that of the United States. While European national supercomputing centers have developed systems capable of both traditional high-performance computing (HPC) simulation and modeling, efforts to fine-tune generative AI and traditional machine learning models are concentrated within these national laboratories rather than among private AI model developers. This disparity exists primarily because leading AI model builders are headquartered in the United States and China. Nevertheless, with national labs spearheading both HPC and generative AI initiatives across Europe, government-funded scientific research is being advanced by experienced researchers who simultaneously allocate compute resources to European enterprises developing AI models and utilizing HPC tools for product design.

Unlike commercial data centers, European national laboratories are not funded by private equity, GPU accelerator manufacturers, or major cloud providers. Consequently, they operate with significantly tighter budgets and reduced compute resources—often two to three orders of magnitude less. Despite these constraints, these facilities are accessing greater compute capacity than ever before, enabling them to pursue scientifically and socially beneficial projects with clear institutional mandates.

This context frames the development of the next-generation Lumi-AI supercomputer, a key component of what industry partners refer to as the “Lumi AI Factory”—a term originally popularized by NVIDIA. Funded by the EuroHPC Joint Undertaking, the system will be deployed at the Center for Scientific Computation (CSC) facilities in Kajaani, Finland. The name “Lumi” translates to “snow on the ground” in Finnish, though the system’s mascot is a white wolf; alternatively, it could have been named “susi,” the Finnish word for wolf.

The original Lumi hybrid supercomputer cluster, unveiled in October 2020, combined CPU-only, CPU-GPU, high-memory CPU, object storage, Lustre parallel file system, and flash storage nodes. It was designed to deliver approximately 552 petaflops of aggregate computing at FP64 precision. The CPU-focused Lumi-C partition, comprising over 200,000 cores based on AMD EPYC processors, was projected to deliver roughly 2 petaflops of FP64 performance, while the GPU-centric Lumi-G partition, built around AMD Instinct processors, was expected to provide approximately 550 petaflops of double-precision floating-point throughput. The broader cluster architecture included 30 petabytes of Ceph object storage, 80 petabytes of Lustre storage, and 7 petabytes of flash storage. All nodes and partitions were interconnected via the “Rosetta” 200 Gb/s Slingshot-10 network, with Hewlett Packard Enterprise serving as the primary system contractor.

EuroHPC allocated a $237 million budget for the initial deployment, equating to approximately $431 per teraflop. The Lumi-C partition was delivered in summer 2021, followed by the Lumi-G GPU partition at the end of 2021. Both became fully operational in June 2022. Even four years after its initial deployment, the Lumi-G partition retained its position as the 11th-ranked system on the June Top500 list, demonstrating sustained competitive relevance.

The Lumi-C partition consists of 1,536 nodes, each equipped with dual 64-core AMD EPYC 7763 “Milan” processors running at 2.45 GHz, totaling precisely 196,860 cores. Ranked 262nd on the June Top500 list, it delivers a peak theoretical performance of 7.29 petaflops and a sustained LINPACK performance of 6.3 petaflops. This represents a computational efficiency of 82.6% within a 1.22-megawatt power envelope, yielding an energy efficiency of 5.18 gigaflops per watt. Node communication within the Lumi-C partition is facilitated by a 200 Gb/s Slingshot-11 interconnect.

The Lumi-G partition comprises 2,978 nodes, each featuring a single 64-core AMD EPYC 7A53 “Trento” custom processor paired with four AMD Instinct MI250X “Aldebaran” GPU accelerators. The Trento chip was originally engineered for the Frontier supercomputer at Oak Ridge National Laboratory, providing unified memory coherence between AMD CPUs and GPUs prior to its commercial availability. In total, the partition houses 2,978 Trento CPUs (190,592 cores) and 11,912 Aldebaran GPUs (2.62 million compute units and 167.7 million stream processors). Interconnected via a 200 Gb/s Slingshot-11 network, the Lumi-G partition achieves a peak FP64 performance of 531.5 petaflops and a LINPACK rating of 379.7 petaflops, reflecting a 71.4% computational efficiency. Operating within a 7.1-megawatt power envelope, it delivers 53.43 gigaflops per watt on LINPACK—more than ten times the power efficiency of the CPU-only partition.

The forthcoming Lumi-AI partition represents a direct upgrade to the Lumi-G architecture. Given the enduring value of AMD’s GPU portfolio and global GPU capacity constraints, the existing Lumi-G partition is expected to remain operational alongside the new system. The Lumi-AI upgrade is projected to deliver ten times the performance on AI workloads and “around 2X” the performance on FP64 HPC tasks.

Surpassing the AI performance of the MI250X will be straightforward. The Aldebaran architecture supported INT4, INT8, and FP16 data formats, but delivered identical throughput across all three, meaning INT4 and INT8 were emulated atop FP16 rather than processed natively—a reasonable compromise at the time. The newer Altair-based MI430X GPUs, designed for both HPC and AI workloads (distinct from the higher-end MI455X optimized for AI inference and training), introduce native FP4 and FP8 processing. This architectural shift yields a multiplicative performance gain as precision decreases: fourfold of the projected tenfold AI performance increase stems directly from native FP4 execution. The remaining 2.5x improvement reflects the raw FP16 performance advantage of the Altair architecture over the Aldebaran platform.

Consequently, the Lumi-AI partition will require significantly fewer nodes, despite the individual accelerators consuming more power and generating greater thermal output. Calculating the target node count requires examining the underlying hardware specifications. The MI250X delivers 47.9 teraflops on vector units and 95.7 teraflops on matrix units. The MI430X achieves a peak of 288 teraflops for FP64 workloads, with AMD specifying that this performance applies equally to both vector and matrix operations. Analyzing the Lumi-G partition’s reported metrics suggests CSC Finland calculated peak performance using vector FP64 throughput: 47.9 teraflops × 4 GPUs per node × 2,978 nodes equals 570.6 petaflops. The discrepancy with the published 531.5 petaflop LINPACK Rpeak remains unexplained.

To achieve a twofold increase in FP64 vector performance for Lumi-AI, only 3,962 MI430X GPUs are required, translating to 991 nodes—a 66.7% reduction in node count. This configuration surpasses the exascale threshold with a peak performance of 1.14 exaflops, a milestone with significant psychological and marketing value. However, power consumption per node will approximately double: the Venice-Altair 1×4 board draws 5 kilowatts compared to the 2.5-kilowatt Trento-Aldebaran 1×4 board. At 991 nodes, the Lumi-AI partition would consume roughly 5 megawatts, representing a 33% reduction in total power draw relative to the Lumi-G partition. Interestingly, scaling this node count to calculate AI performance yields only an eightfold increase. To achieve the stated tenfold AI performance boost at 4-bit precision, 1,240 nodes equipped with four MI430X GPUs each would be necessary. This discrepancy suggests CSC Finland may ultimately secure approximately 25% more FP64 vector performance than initially communicated.

Based on these calculations, the most likely configuration consists of 1,240 nodes housing 4,960 MI430X GPUs. This represents a 58.4% reduction in node count compared to the GPU partition and an estimated system power draw of 6.2 megawatts. Primary contractor Bull and CSC Finland have explicitly stated that the Lumi-AI partition will deliver 10X the AI performance and HPC performance would be “about 2X.” The inherent imprecision in these projections remains puzzling, particularly for institutions that traditionally prioritize exactitude.

Leveraging the 256-core AMD EPYC 9006 “Venice” processor, the Lumi-AI partition will feature 317,440 CPU cores. This substantial increase provides ample compute resources to orchestrate the GPU accelerators and represents a 66.7% growth in core density compared to the existing Lumi-G partition.

With a projected budget of €387.8 million for the Lumi-AI partition—equivalent to $445.1 million—and a peak FP64 performance of 1.43 exaflops, the system will cost approximately $312 per teraflop at peak. This represents a 27.6% reduction in cost per teraflop compared to the $431 per teraflop peak cost of the original Lumi cluster deployed in 2021. It remains unknown whether the existing CPU-only partition and associated storage clusters will be integrated into or phased out alongside the new deployment.