SSN_-_Amendment_2_4.pdf

United States Department of the Interior

INTERIOR BUSINESS CENTER

Washington, DC 20240

Computational Infrastructure — GPU Server Node

Amendment 2 extends the closing date and time to July 24, 2026 at 5 P.M. EDT and provides answers to questions, as follows:

1.CPU requirements – what are you preferred CPU architecture, socket count, core count, generation and any minimum core/clocl requirement for data preprocessing and model training?

The preferred architecture is an enterprise-class dual-socket server utilizing the latest available generation of Intel Xeon or AMD EPYC processors. There is no preference between Intel and AMD provided equivalent performance is achieved.

The preferred configuration should include 128–192 total physical CPU cores (64–96 cores per socket), with high memory bandwidth and PCIe Gen5 support to maximize GPU performance and data throughput. The platform should be optimized for simultaneous execution of laboratory automation workflows, high-content image processing, AI inference, model development, virtualization, and data preprocessing.

Note that while L40s (GPUs) aren’t preferred, they would be acceptable. The RTX 6000s would make it compatible with the cluster for SOM so there’d be no issues with both running the same software.

2. CPU RAM – how much ECC DDR5 system ram is required (RFI only mentions GPU memory)?

The preferred configuration is 1 TB ECC DDR5 RAM, expandable to 2 TB or greater. The system memory should support concurrent execution of AI/ML workflows, large imaging datasets, virtualization, database services, laboratory information management systems (LIMS), and multimodal data integration.

3. Storage – what’s the preferred media type/performance (all NVMe SSD, hybrid SSD/HDD or high capacity SATA/SAS HDD) probably would have to be?

The preferred storage architecture is a tiered enterprise storage solution consisting of:

Primary storage: Enterprise NVMe SSDs for operating system, applications, active AI workloads, databases, and high-performance image processing.

Secondary storage: High-capacity enterprise SSD or SAS storage for research datasets and imaging repositories.

The system should provide 100 TB usable capacity at deployment while supporting expansion to 250–500 TB without major architectural changes. Storage should be optimized for high I/O performance and concurrent access by multiple users and AI workflows.

4. Do you need software RAID?

No preference for software RAID. The preferred solution is an enterprise-grade hardware RAID controller (or equivalent enterprise storage architecture) supporting RAID 6 or comparable fault-tolerant configurations, hot-spare capability, and enterprise-level data protection.

5. Networking – what’s their networking interface and speed needs?

The preferred networking configuration is:

Dual redundant 100 Gb Ethernet interfaces for high-performance data movement.

Additional 25 Gb and 10 Gb Ethernet interfaces for laboratory instrument connectivity, management, and institutional network integration.

The architecture should support future integration with high-performance computing resources, cloud environments, robotics platforms, imaging systems, and AI infrastructure.

6. System size – do they have constraints? 2U/4U/5U etc There are no specific rack unit constraints. The preferred solution is an enterprise rack-mounted system (4U–8U or larger if required) that provides adequate cooling, PCIe expansion, GPU capacity, and future upgradeability. System architecture and expandability are more important than minimizing rack footprint.

7. Do they need the Nvidia software NVAIE?

Yes. Support for NVIDIA AI Enterprise (NVAIE) or equivalent enterprise AI software is preferred. The platform should support enterprise AI frameworks for model development, inference, containerized AI workflows, computer vision, high-content image analysis, digital twin applications, and autonomous laboratory operations. Compatibility with NVIDIA CUDA, TensorRT, Kubernetes, and future AI software ecosystems is highly desirable.