The first module of the JUPITER exascale supercomputer, named JEDI, has taken the top spot on the Green500 list, which ranks the most energy-efficient supercomputers in the world. JUPITER was installed in April by the German-French consortium and shares the same hardware as the booster module JUPITER currently being built at Forschungszentrum Jülich.
This supercomputer has become a pioneer in the field. This first module is capable of performing 72,000 billion floating point operations per second per watt, while the previous leader reached 65,000 billion.
The key factor in the extraordinary efficiency of the module is the use of graphics processing units (GPU) and the ability to optimize scientific applications to perform calculations on them. Currently, almost all leading systems on the Green500 list heavily rely on GPUs, which are designed to perform calculations with much higher energy efficiency than conventional central processing units (CPUs).
The JEDI development system is one of the first in the world to use the latest generation of NVIDIA accelerators: the NVIDIA GH200 Grace Hopper superchip, which combines the NVIDIA Hopper GPU and the NVIDIA Grace CPU into a single module. Based on the latest BullSequana XH3000 architecture from Eviden, the system includes their highly efficient Direct Liquid Cooling hot water cooling system, which requires much less energy than conventional air cooling and allows for the reuse of generated heat afterwards.
The precursor to JUPITER, JEDI, already features the same equipment as the subsequent JUPITER booster module. Under the JUPITER Research and Early Access Program (JUREAP), scientists can access the hardware in an early development phase to optimize their codes with the support of experts from the Jülich Supercomputing Centre.
JUPITER will be the first European supercomputer to surpass the exaflop threshold, corresponding to a quintillion (1 followed by 18 zeros) floating-point operations per second. The final system will be installed in stages in the second half of this year and initially made available to scientific users under the early access program before becoming operational for general users in early 2025.
The immense computing power of JUPITER will help push the boundaries of scientific simulations and train large AI models. Additionally, it utilizes the dynamic modular system architecture (dMSA) developed by ParTec and the Jülich Supercomputing Centre. The installed JUPITER booster module will consist of approximately 125 BullSequana XH3000 racks and around 24,000 NVIDIA GH200 Superchips interconnected via the NVIDIA Quantum-2 InfiniBand network. For 8-bit calculations, common in AI model training, the computing power will far exceed 70 exaflops, making JUPITER the fastest AI computer in the world today.
Estimated energy requirements for JUPITER will average around 11 megawatts. Other measures will contribute to even more sustainable energy use. The modular data center housing JUPITER is designed to extract heat generated during cooling and use it to heat buildings on the Forschungszentrum Jülich campus.
All JUPITER hardware and software components will be installed and managed using the exclusive JUPITER management stack, a combination of ParaStation Modulo (ParTec), SMC xScale (Eviden), and software components from JSC.
The JUPITER JEDI development system is much smaller than the final exascale computer. It consists of a single rack of the latest BullSequana XH3000 series, currently containing 24 individual computers, known as compute nodes. These are interconnected via four NVIDIA Quantum-2 InfiniBand switches and will be complemented with an additional 24 computing nodes throughout May.
During measurements for the Green500 ranking of the most energy-efficient supercomputers, the JEDI system achieved a computing power of 4.5 quadrillion floating-point operations per second, or 4.5 petaflops, with an average consumption of 66 kilowatts. During optimized operation, consumption was reduced to 52 kilowatts.