Overview
The Katea Computing Platform (KCP) is Tecnalia's first supercomputing center, strategically located at Mikeletegi 2 in Donostia, Spain.
Katea KCP is a state‑of‑the‑art HPC system designed for large‑scale scientific computing, artificial‑intelligence research, and data‑intensive workloads. It combines a heterogeneous mix of high‑core‑count CPUs, the latest NVIDIA GPUs, a fast Lustre file system, and a 200 Gb/s InfiniBand fabric to deliver near‑peak performance on a wide variety of applications.
Hardware Specifications
| Component | Details |
|---|---|
| Platform name | Katea KCP |
| Operating system | Red Hat Enterprise Linux 9 |
| Interconnect | Mellanox InfiniBand 200 Gb/s (HDR) |
| Storage (Lustre) | 230 TB high‑throughput parallel file system |
| Node count | 19 compute nodes |
| Total CPU cores | 1 108 cores |
| CPU types | • AMD EPYC 9374F (64 cores per socket) • Intel Xeon Gold 6444Y (48 cores per socket) |
| CPU performance per node | ~8 800 GFLOP/s |
| Maximum CPU performance | 120 448 GFLOP/s (≈ 120.4 TFLOP/s) |
| GPU types | • NVIDIA H100 80 GB • NVIDIA A100 80 GB |
| GPU performance (FP32) | 690 TFLOP/s |
| GPU performance (FP64) | 350 TFLOP/s |
| Main memory | 9.7 TB total (mixed configurations) |
| Memory per node | • 10 × 388 GB DDR4 • 8 × 512 GB DDR4 • 1 × 1.5 TB DDR4 |
| Local NVMe storage | • 480 GB (CPU‑only nodes) • 3 TB (GPU‑enabled nodes) |
CPU Details
- AMD EPYC 9374F – 64 cores, 2.3 GHz base clock, 128 MB L3 cache, supporting AVX2/AVX‑512.
- Intel Xeon Gold 6444Y – 48 cores, 2.5 GHz base clock, 66 MB L3 cache, AVX‑512 support.
The heterogeneous mix allows users to select the most appropriate architecture for their workload, balancing raw core count, memory bandwidth, and instruction‑set features.
GPU Details
- NVIDIA H100 80 GB – Hopper architecture, ~3 TB/s memory bandwidth, Tensor‑Float‑32 (TF32) and FP64 acceleration.
- NVIDIA A100 80 GB – Ampere architecture, ~2 TB/s memory bandwidth, versatile for both AI and HPC kernels.
Together, the GPUs deliver 690 TFLOP/s (FP32) and 350 TFLOP/s (FP64) of peak floating‑point performance, making Katea KCP one of the most powerful AI‑ready platforms in Spain.
Network & Storage
- InfiniBand HDR 200 Gb/s ensures low‑latency, high‑throughput communication for distributed‑memory simulations and large‑scale AI training.
- Lustre (230 TB) provides up to ~100 GB/s aggregate bandwidth, optimized for both sequential I/O (e.g., checkpointing) and random I/O (e.g., database queries).
Applications
The platform’s balanced architecture enables high sustained efficiencies across a diverse workload spectrum:
- Scientific Research: KCP can support a wide range of scientific research, including physics simulations, climate modeling, and genomics.
- Machine Learning and AI: The powerful GPU configuration makes KCP ideal for training and deploying complex machine learning models and deep neural networks.
- Data Analytics: The large RAM and storage capacity enable efficient handling of big data analytics, data mining, and complex database operations.
- Engineering Simulations: KCP can perform detailed simulations for various engineering applications, such as aerodynamics, structural analysis, and materials science.