Germany - Computer equipment and accessories

Description

The Institutes of Engineering Mathematics and Technology of Autonomous Systems intend to procure a GPU computing cluster within the DTEC.bw projects SeRANIS and MORE for the development and research of deep learning and optimization algorithms in space debris removal as well as environment recognition and behavior decision of autonomous road vehicles. Since CUDA-based software and long-standing know-how with regard to Nvidia's architecture are already available, UniBw M urgently requires a computing cluster consisting of GPU modules that enables the execution and development of CUDA-based software solutions. In addition, UniBw M requires a storage system that also serves as a data interface to the GPU computing cluster.
The Institutes of Engineering Mathematics and Technology of Autonomous Systems intend to procure a GPU computing cluster within the DTEC.bw projects SeRANIS and MORE for the development and research of deep learning and optimization algorithms in space debris removal as well as environment recognition and behavior decision of autonomous road vehicles. Since CUDA-based software and long-standing know-how with regard to the architecture are already available from Nvidia, UniBw M urgently needs a computing cluster consisting of GPU modules that enable the execution and development of CUDA-based software solutions. In addition, the UniBw M requires a storage system that also serves as a data interface to the GPU computing cluster.GPU computing cluster:Central requirements for the GPU computing cluster, consisting of four DGX-A-100-80 modules, are for each module:1. The basis for the computing cluster is compatibility of the drivers and the microarchitecture of the GPU with CUDA-based software and corresponding libraries.2.Compute Capability 8.0 and Shared Memory per Streaming Machine (SM) of at least 164 KB.3 DRAM availability of at least 80 GB per GPU, with a bandwidth of 2 000 GB/s or higher. In addition, DRAM available per module must be at least 640 GB.4. Double Precision support of available Tensor cores, for highly accurate computational results essential in scientific applications.5. Computational performance in Single Precision 19.5 Teraflops (TF) and Tensor Float (Single Precision) 312 TF for sparse matrices, as well as Tensor Float (Double Precision) at 19.5 TF.6.Dual CPU-on-board system with 64 cores each and a base clock of 2.25 GHz, 2x 1.92 TB M.2 NVME memory for the operating system (as RAID-1) as well as 8x TB U.2 NVME memory for intermediate storage of calculation results and 2 TB RAM.7. Complete networking of the individual modules with a bandwidth of at least 100 Gbit/s using Infiniband and a corresponding Infiniband switch. The Infiniband technology already realizes the connection between the GPUs on the DGX boards by means of the NVLink version 3.0. A corresponding link to the management system with the same standard is therefore required to maintain conformity in bandwidth.8. hardware support of 5 years. This must include a hardware warranty with on-site replacement service for critical components.9.. software support for 5 years and access to the appropriate Nvidia developer repositories,tools and libraries and associated documentation.Storage System:Key Storage System Requirements:1. SSD-based storage volume of at least 550 TB (gross) distributed across 3 servers,2. HDD-based storage volume of at least 1 792 TB (gross) distributed across 4 servers Networking,1. Broadband and low-latency networking of the individual GPU models of all DGX servers with at least 100 Gbit/s Infiniband HDR,2. Connection of the DGX servers to the storage and CPU servers with at least 200 Gbit/s Infiniband HDR,3. All servers must have a separate network interface with at least 10 Gbit/s for the connection to the host data networkInstallation, maintenance, support, administration Key requirements:1. Installation and commissioning of the entire system on site.2.Maintenance of the software standard through regular updates (including operating system, firmware).3. The service provider for maintenance / operation must be German-speaking and have many years of expertise in the support and maintenance of comparable systems.4. Availability and accessibility of support at the latest on the next working day.5. Management system for the scheduling and deployment of jobs on the DGX-A100. Further details, in particular regarding the headnode and CPU server, can be found in the service description.

Opportunity closing date

20 August 2021

Value of contract

to be confirmed