Bull has been selected over Hewlett Packard Enterprise as the primary contractor for the next-generation Lumi-AI partition at CSC Finland in Kajaani. Funded by the EuroHPC Joint Undertaking with contributions from Finland, Czechia, Denmark, Estonia, Norway, Poland and the European Union, the system is expected to be installed in the second half of 2027 in a new data center. Lumi-AI is described as an upgrade to the existing Lumi-G GPU partition, whose AMD Instinct MI250X systems provide 379.7 petaflops of sustained FP64 Linpack performance and 531.5 petaflops of theoretical peak performance. Bull and CSC Finland say the new partition will deliver 10 times higher AI performance and about twice the FP64 HPC performance. The article attributes much of the AI gain to AMD’s Altair-based MI430X GPUs, which support native FP4 and FP8 processing; the shift from FP16 to FP4 is estimated to provide a fourfold multiplier, with another 2.5 times attributed to higher raw FP16 performance. The author estimates that Lumi-AI could use either 991 nodes with 3,962 GPUs for roughly twice the FP64 vector performance, or 1,240 nodes with 4,960 GPUs to reach the stated 10-fold AI increase at four-bit precision. The latter estimate would represent a 58.4% reduction in GPU-partition node count, around 6.2 megawatts of system power, and up to 1.43 exaflops of FP64 peak performance, although the exact configuration was not disclosed. Each node is expected to use a 256-core AMD Venice Epyc 9006 processor, producing an estimated 317,440 CPU cores. The contract is valued at €387.8 million, or about $445.1 million in the article’s conversion, implying approximately $312 per peak FP64 teraflop compared with $431 for the original Lumi cluster. The article says Bull’s BXI protocol is being adapted for Ultra Ethernet Consortium hardware, potentially giving BullSequana XH3500 customers more interconnect options. The wider Lumi AI factory will also include the Lumi-IQ quantum system from IQM, with a planned 150-physical-qubit Halocene H4 installation in 2027, upgrades in 2028, and a Halocene H5 system planned for 2029 with at least nine logical qubits.
