The end customer in this project is a German manufacturer of specialized high-performance computing (HPC) systems with a focus on FPGA-accelerated computing. The company develops modular parallel computers that are used in areas such as cryptography, simulation, bioinformatics, and algorithmic research.
With this technology, field programmable gate arrays (FPGAs) take on particularly computationally intensive tasks directly at the hardware level. This allows certain algorithms to be executed much faster and more energy-efficiently than on conventional CPU or GPU architectures.
To develop and validate such systems, the customer requires a robust, scalable server platform that reliably supports the parallel use of CPU, GPU, and FPGA workloads.
Project period: Q2/2025
Project volume: HPC systems in the six-figure range
Table of Contents
Project Description
The aim of the project was to set up a new HPC cluster platform for performing highly parallel calculations with GPU and FPGA support. The customer needed a scalable server system that was suitable for both classic software simulations and hardware-accelerated computing operations.
The focus was on integrating a powerful, PCIe-optimized server infrastructure that could be flexibly equipped with different FPGA and GPU cards. In addition to high computing power, energy efficiency, reliability, and ease of maintenance were key factors.
The systems were to be integrated into 19" racks on a turnkey basis and delivered with complete cabling, power distribution, and network connectivity. The entire cluster had to be tested under Linux and prepared for immediate productive use.
Key Requirements
- High parallel computing power for CPU, GPU, and FPGA-based applications
- PCIe preparation for modular FPGA integration
- Large memory capacity and expandability for simulation workloads
- Fast NVMe storage with redundant data backup
- Redundant 10 Gigabit network connection for data and management network
- Turnkey rack integration including KVM console, PDU, and cabling
- Complete functional testing under Linux for reliable continuous operation
Project Implementation
- Server platform
21× Supermicro SuperServer 420GP-TNR
→ Rack-optimized 4U platform with high computing density, redundant power supply, and cooling designed for GPU and FPGA expansions. Ideal for hybrid HPC workloads that combine CPU, GPU, and FPGA acceleration. - Processors (2 per server)
42× Intel Xeon Silver 4310 (12 cores / 24 threads, 2.10 GHz base, 3.30 GHz turbo, Ice Lake-SP, 120 W TDP)
→ Dual-socket setup with a total of 24 physical and 48 logical cores per node. Provides a stable and energy-efficient foundation for parallel HPC workloads and control processes of FPGA clusters. - Memory (6 × 64 GB per server)
126× Samsung DDR4-3200 ECC reg. DR (total: 384 GB per node, effective 2666 MHz)
→ High memory capacity with error correction for stable long runs and simulations. 18 DIMM slots remain free for future expansion. - System storage (1 per server)
21× Samsung PM9A3 M.2 NVMe SSD (1.92 TB, PCIe Gen4 x4, up to 5,500 MB/s read, 2,000 MB/s write)
→ Fast, fail-safe boot drive with hardware encryption and power loss protection – ideal for Linux-based cluster operating systems. - Network (1 module per server)
21× Supermicro AOC-ATGC-i2TM AIOM (2× 10 GbE RJ45)
→ Dual 10 Gigabit interface via OCP3 module for redundant connection to data and management networks. - Assembly & testing
Pre-assembly and certification according to ISO 9001:2015 / ESD IEC 61340-5-1
PCIe power cable per dual x16 slot pre-assembled for FPGA slots
Functional test and burn-in under Linux (Red Hat derivative)
→ Securely tested delivery as a turnkey HPC system, separately packaged for transport and integration at the end customer's site. - Service level
3-year parts warranty, SLA 1/3 (5×9) incl. MUSTANG® systems ticket system
Result
With the delivered solution, the customer received a fully integrated, turnkey HPC cluster consisting of 21 powerful compute nodes based on the Supermicro 420GP-TNR platform. Thanks to the preconfigured rack integration, including network, power distribution, and KVM infrastructure, the entire system could be put into operation immediately after delivery.
The combination of dual Xeon architecture, generous ECC memory, and NVMe storage creates a stable foundation for parallel FPGA and GPU calculations. Thanks to the pre-assembled PCIe power cables and modular AIOM network connection, the system is optimally prepared for the use of different accelerator cards and future expansions.
The integrated rack infrastructure with a dedicated data and management network ensures high operational reliability, clear separation of workloads, and efficient maintenance processes. Thanks to comprehensive functional testing under Linux, the cluster environment could be transferred directly into productive research and development operation.
The result is a scalable, energy-efficient, and service-friendly HPC platform that ideally complements the customer's FPGA-based special-purpose computers and opens up new possibilities for computationally intensive simulations, algorithmic analyses, and hardware evaluations.