Komondor is currently in a pilot phase, altogether it has the capacity of 5 petaflops.
Regarding its architecture, approximately 80% of this capacity is provided by the GPU-accelerated partition, but the system include a separate, purely CPU-based partition, as well as dedicated artificial intelligence (AI) and Big Data partitions. The system is equipped with a multi-level storage system of more than 10 petabytes.
This supercomputer utilizes an extremely advanced and energy-efficient technology, and thus has a power demand less than 400 kW despite its huge computing performance.
The machine use an especially environmentally-friendly solution for cooling – the so-called hot water cooling technology –, which enables the reuse of the waste heat generated by the machine, as well as a highly energy-efficient passive cooling.
- “CPU only” partition
- 184 nodes, each with a 2 x 64-core AMD EPYC™ 7763 (Milan) CPU and 256 GB RAM (a total of 17920 cores), 200Gb/s Slingshot interconnect
- Performance: 0,9+ petaflops (Rmax)
- “Accelerated (GPU)” partition
- 58 nodes, each with a 1 x 64-core AMD EPYC™ 7763 (Milan) CPU and 256 GB RAM, and 4 x NVIDIA A100 GPUs (a total of 200 GPUs), 2x200Gb/s Slingshot interconnect
- Performance: 4,6+ petaflops (Rmax)
- “Artificial Intelligence (AI)” partition
- 4 nodes (HPE Apollo 6500 Gen10 Plus), each with a 2 x 64-core AMD EPYC™ 7763 (Milan) CPU and 512 GB RAM and 8 NVIDIA A100 GPUs (a total of 16 GPUs), 2x200Gb/s Slingshot interconnect
- Performance: 0,6+ petaflops (Rmax)
- “Big Data (Data Analytics)” partition
- 1 node (SMP/NUMA), 16 x 18-core Intel® Xeon® Gold 16 CPU (a total of 288 cores) and 12 TB RAM, 2 TB SSD, 2x200Gb/s Slingshot interconnect
- Performance: 30 teraflops (Rmax)
- 3-level integrated data storage with a 400 TB ultrafast storage unit for the workflow area (300 Gbyte/s), 1,5 PB fast storage unit for short-term storage, and a10 PB long-term archiving system.