Skip to content

Architecture of the PERUN Supercomputer

Overview

The PERUN Supercomputer is the core high‑performance system of the Supercomputing Center at the Technical University of Košice.
It is engineered for large‑scale, high‑throughput scientific and industrial workloads.


1. System Architecture

PERUN consists of two complementary compute partitions designed to meet diverse HPC and AI workloads:

  • PERUN Universal (CPU Partition) – general‑purpose HPC and memory‑intensive simulations
  • PERUN AI (GPU Partition) – GPU‑accelerated AI, ML, and massively parallel workloads

2. PERUN Universal – CPU Partition

Best For

Scientific simulations, large memory workloads, classical HPC, parallel processing.

2.1 Hardware Specification

Component Specification
Nodes 32 × HPE Cray XD2000 (XD225v)
CPU 2 × AMD EPYC 9745
Cores per Node 256
Memory per Node 1,536 GB DDR5 ECC
Network • 2 × 100 Gb/s Ethernet
• 1 × 200 Gb/s NDR200 InfiniBand

2.2 Advantages

  • Excellent parallel performance
  • Ideal for traditional HPC codes
  • Outstanding memory bandwidth and capacity per node

3. PERUN AI – GPU‑Accelerated Partition

Best For

Deep learning, AI training, large‑scale simulations, GPU computing, data analytics.

3.1 Hardware Specification

Component Specification
Nodes 26 × HPE ProLiant Compute XD685
CPU 2 × AMD EPYC 9535
Cores per Node 128
GPU 8 × NVIDIA H200
GPU Memory 141 GB HBM3e per GPU
System Memory 2,304 GB DDR5 ECC
Network • 2 × 100 Gb/s Ethernet
• 4 × 400 Gb/s NDR InfiniBand
Interconnects • GPU ↔ GPU: 900 GB/s
• GPU ↔ CPU: 128 GB/s
• NVLink internal GPU fabric

3.2 Advantages

  • Extreme performance for deep learning and large AI models
  • High‑throughput NVIDIA Hopper architecture
  • Optimized for multi‑GPU workloads and parallel data pipelines

4. Performance, Flexibility & Connectivity

PERUN’s hybrid CPU+GPU architecture enables optimal resource matching for every type of workload.

4.1 System Performance

Metric Value
Total Performance (Rmax) 10.7 PFlops

4.2 Connectivity & Storage

Component Description
Network High‑speed InfiniBand fabric
Storage Integrated HPC PERUN data storage system

Result

The infrastructure ensures high‑speed access, efficient processing of large datasets, and maximum utilization of PERUN’s computational power.


If you'd like, I can also prepare a dark‑mode styled version, add images, or integrate this into your full PERUN documentation.