Attachment 1 - SOW AI_HPC Server.pdf

Statement of Work (SOW) for

AI/HPC Server at the

704th Test Support Squadron

8 January 2026

1.0 PURPOSE

The 704th Test Group (704 TG) requires one (1) GPU-accelerated rackmount server system to support mission-critical Artificial Intelligence (AI) training, deep learning, and High-Performance Computing (HPC) workloads. This document defines the minimum technical and service requirements for the system.

2.0 SCOPE

The vendor shall provide one (1) fully configured Supermicro GPU SuperServer SYS-522GA-NRT, or a fully equivalent system, that meets or exceeds the following component and performance specifications.

2.1 Technical Specifications

2.1.1 System Platform

• Supermicro SYS-522GA-NRT 5U rackmount GPU server (or equivalent)

• Designed for: AI / Deep Learning, 3D Rendering, High Performance Computing.

2.1.2 Processor

• (Quantity: 2) Enterprise-class server processors equivalent to or better than:

Intel® Xeon® 6952P, 96-core, 2.10 GHz, 480 MB cache, ~400 W TDP

• Processors must meet or exceed the following minimums:

• Minimum 192 physical cores total

• Support for DDR5-6400 ECC RDIMM

• Enterprise virtualization support (VT-x/AMD-V, SR-IOV, IOMMU equivalent)

• Sustained performance equivalent to or greater than Xeon 6952P class CPUs

• Validated for multi-GPU server platforms

• Vendor shall provide documentation demonstrating performance equivalency if alternate processors are proposed.

2.1.3 Memory

• (Quantity: 16) 32 GB DDR5-6400 ECC RDIMM modules

• Total system memory: 512 GB DDR5 ECC

• Must support ECC, memory sparing, and advanced RAS features

2.1.4 Storage

• (Quantity: 1) 480 GB M.2 NVMe SED solid-state drive

• Model class: PE9010 or equivalent

• Used for operating system / boot volume

• System must support hot-swap NVMe expansion bays

2.1.5 GPU / Accelerator

• (Quantity: 4) Data-center or enterprise-class GPUs equivalent to or better than:

NVIDIA RTX PRO 6000 Blackwell Server Edition (96 GB VRAM class)

• Each GPU must meet or exceed the following minimums:

• Minimum 96 GB VRAM per GPU

• Designed for continuous server operation

• Support for virtualization (vGPU or equivalent GPU partitioning technology)

• PCIe Gen4 or Gen5 interface

• Workload suitability for AI Training and HPC

• Thermal design compatible with rackmount servers

Vendor shall provide specification sheets and benchmark documentation demonstrating equivalent or superior performance and memory capacity if alternate GPUs are proposed.

2.1.6 Performance and Acceptance Criteria

Proposed equivalent CPUs or GPUs must:

• Meet or exceed the baseline system’s performance in at least one of the following:

o AI/ML compute benchmarks o HPC floating-point throughput o GPU memory capacity and bandwidth

• Be supported by the server OEM (not white-box unsupported cards)

• Be covered under the system warranty

The purchaser reserves the right to approve or reject substitutions based on documented performance data.

2.1.7 Networking

• Integrated (Quantity: 2) 10 GbE RJ-45 Ethernet ports

• Must support iSCSI, NFS, SMB, and virtualization networking use cases

2.1.8 Expansion and Architecture

• PCIe Gen5 backplane and riser architecture

• Capable of supporting multi-GPU compute workloads and high-throughput NVMe storage

2.1.9 Power and Cooling

• Redundant, hot-swappable power supplies

• Heavy-duty high-airflow cooling design suitable for continuous GPU operation

• System must support full thermal load of dual CPUs and four 600 W GPUs

2.1.10 Rack Compatibility

• 5U rackmount chassis

• Compatible with standard 19-inch, 24” x 40” server racks

• Must include rack rails and mounting hardware

2.2 Included Hardware and Accessories

• (Quantity: 1) Supermicro rail kit MCP-290-00057-0N (included)

• All internal power cables, GPU power distribution, and airflow shrouds required for the configured system

2.3 Service Requirements

2.3.1 Warranty and Support

• 3 years parts and labor warranty

• 1 year cross-shipment replacement coverage

• Drive-retention in the case warranty replacement.

• Technical telephone support and hardware replacement included

2.3.2 Installation and Validation

• System shall be delivered fully assembled and tested.

• Pre-installed OS and licensing is not required.

2.3.3 Documentation

• Full system configuration summary

• Manufacturer datasheets

• Warranty certificates

• Support contact information

3.0 MILESTONES

The contractor shall deliver 1 fully configured GPU server system no later than 60 days after receipt of order (ARO).

4.0 DELIVERABLES

The contractor shall deliver the following to 1644 Vandergrift Road, Building 1265, Holloman Air Force Base, New Mexico 88330. Attn: Vaden Gilloth, vaden.gilloth.1@us.af.mil, 575-572-3825

• 1 × Supermicro SYS-522GA-NRT GPU server system (or equivalent and configured as specified)

• Installed CPUs, memory, GPUs, and storage

• Rail kit and rack mounting hardware

• All manufacturer documentation and warranty materials

• Configuration and validation report mailto:vaden.gilloth.1@us.af.mil