EuroHPC Joint Undertaking has awarded Eviden's Bull the contract for the next-generation Lumi-AI supercomputer, beating incumbent Hewlett Packard Enterprise on a €387.8 million ($445.1 million) deal. The system will be installed at CSC Finland's new datacenter in Kajaani in the second half of 2027.

The contract delivers 1.43 exaflops of peak FP64 performance at a cost of $312 per teraflops — 27.6 percent lower than the $431 per teraflops the original Lumi cluster cost when it was installed in 2021.

CSC Finland and Bull stated the Lumi-AI partition will deliver 10X the performance on AI workloads and "around 2X" the performance on FP64 HPC workloads. Both the contractor and the hosting center declined to give precise HPC figures — an imprecision that is counterintuitive for organizations that live by precision.

What does the new partition replace?

The Lumi-AI upgrade effectively replaces the Lumi-G GPU partition, which reached full operation in June 2022 under an HPE-led, $237 million EuroHPC contract. The existing partition runs 2,978 nodes, each pairing a 64-core AMD "Trento" Epyc 7A53 with four "Aldebaran" MI250X GPUs: 190,592 CPU cores and 11,912 GPUs in total.

Lumi-G delivers 531.5 petaflops peak FP64 and 379.7 petaflops sustained on Linpack — 71.4 percent efficiency — within a 7.1 megawatt envelope, or 53.43 gigaflops per watt. It ranked number 11 on the June Top500 list at four years of age.

The CPU-only Lumi-C partition, ranked 262 in June, holds 1,536 dual-socket nodes with 196,860 "Milan" Epyc 7763 cores, 7.29 petaflops peak and 6.3 petaflops Linpack — 82.6 percent efficiency at 1.22 megawatts.

How much of the 10X comes from the silicon?

The MI430X packages — built on AMD's "Altair" GPUs aimed at both HPC and AI, unlike the inference-focused MI455X — process FP4 and FP8 natively, where the MI250X grafted INT4 and INT8 on top of FP16 with no additional throughput. That data-format shift from FP16 down to FP4 accounts for 4X of the claimed 10X AI gain. The remaining 2.5X comes from raw FP16 performance gains comparing Altair against Aldebaran.

On FP64, the MI430X delivers 288 teraflops on both vector and matrix units, against 47.9 teraflops vector and 95.7 teraflops matrix for the MI250X.

Extrapolating from those figures: doubling FP64 vector performance requires roughly 3,962 MI430X GPUs across 991 nodes — a 66.7 percent node-count reduction that still breaks the exaflops barrier at 1.14 exaflops peak. But that configuration only yields 8X on 4-bit AI workloads. Reaching the full 10X requires approximately 1,240 nodes and 4,960 MI430X GPUs, consuming about 6.2 megawatts, which suggests CSC Finland may get 25 percent more FP64 vector performance than it is advertising.

Each Venice-Altair 1x4 board burns about 5 kilowatts, double the 2.5 kilowatts of the Trento-Aldebaran boards. With 256-core "Venice" Epyc 9006 processors, the partition would carry 317,440 CPU cores — 66.7 percent more than Lumi-G — to orchestrate the GPUs.

Why Bull and not HPE?

EuroHPC prefers European suppliers where possible, which favored Bull. Interconnect strategy played a role as well: while HPE's Slingshot is an Ethernet variant that could potentially adhere to Ultra Ethernet Consortium specifications, Bull's BXI protocol is being adapted to run natively on UEC hardware. That gives BullSequana XH3500 customers a choice of switching fabric for scale-out clustering.

Funding comes from the European Union's EuroHPC JU plus Finland, Czechia, Denmark, Estonia, Norway and Poland. The individual contributions were not disclosed. It is not known whether the Lumi-C partition and storage clusters will also be upgraded under the contract, though that is presumed.

What about the quantum component?

The Lumi AI factory will include a quantum computing element, Lumi-IQ, built by IQM Quantum Computers. The Finnish company has raised €600 million in venture capital, holds a $1.8 billion valuation, and is preparing to go public.

IQM will install a Halocene H4 system at CSC Finland in 2027 with 150 physical qubits supporting 5 logical qubits. Error-correction and logical-qubit reliability upgrades follow in 2028. In 2029, CSC Finland receives a Halocene H5 with at least 9 logical qubits — implying roughly 270 physical qubits.