Hardware Evaluation & Budget Brief: NVIDIA-Powered Mini PC Workstation

Document ID: SEN-4

Date: August 19, 2026

Audience: Sentinel Integrations (Michael Morgan)

Objective: Identify and budget a high-density, low-power SFF/Mini PC with dedicated NVIDIA CUDA acceleration to replace aging NUC hardware for local sovereign agent inference.

1. Executive Summary & Selection Criteria

For sovereign local AI execution, standard CPU-only NUCs are bounded by DDR5 memory bus bandwidth (~5–8 t/s generation). Adding dedicated NVIDIA CUDA cores and Tensor Cores unlocks:

• Zero-Latency Tool & Routing Checks: 80–120+ t/s on 7B–12B models.
• Local Embedding & Reranking: Millisecond latency on mxbai-embed-large and bge-reranker.
• Local Vision & Multimodal: Fast execution of OCR, layout parsing, and image analysis.

Target Specifications:

• VRAM: Minimum 8GB–16GB dedicated (or unified on Jetson).
• RAM: 64GB+ DDR5 expandable for large system context buffers.
• Form Factor: Compact SFF/Mini PC (< 3.0 Liters) with low idle power (< 25W).
• Budget Target: $1,200 – $2,200 total capital outlay.

2. Top Hardware Options & Comparison Matrix

Option A: ASUS ROG NUC (RTX 4070 Laptop GPU)

• Processor: Intel Core Ultra 9 185H (16 Cores, 22 Threads, NPU built-in)
• GPU: NVIDIA GeForce RTX 4070 Mobile (8GB GDDR6, 140W max TGP)
• RAM: Up to 64GB DDR5 5600MHz (2x SO-DIMM)
• Storage: 3x M.2 2280 PCIe 4.0 slots
• Form Factor: 2.5 Liters (270 x 180 x 50 mm)
• Connectivity: 2.5GbE LAN, Wi-Fi 6E/7, Thunderbolt 4
• Estimated Cost: $1,699 – $1,999 (Barebones / Configured)
• Verdict: Direct successor to the Intel NUC Enthusiast line (ASUS acquired the NUC brand from Intel). Enterprise build quality, whisper-quiet thermal profile.

Option B: Minisforum AtomMan G7 Ti / G7 Pt

• Processor: Intel Core i9-14900HX (24 Cores, 32 Threads)
• GPU: NVIDIA GeForce RTX 4070 Mobile (8GB GDDR6, 140W TGP)
• RAM: Up to 96GB DDR5 5600MHz
• Storage: 2x M.2 PCIe 4.0 NVMe
• Form Factor: Vertical desktop stand (ColdWave cooling system)
• Connectivity: 2.5GbE, Wi-Fi 7, USB4
• Estimated Cost: $1,299 – $1,599
• Verdict: Highest raw CPU core density (24 cores) and highest RAM expandability (96GB), best price-to-performance ratio.

Option C: Minisforum MS-01 + Low-Profile Workstation GPU

• Processor: Intel Core i9-13900H / i9-12900H
• GPU Slot: Built-in PCIe 4.0 x16 (x8 electrical) low-profile slot
• Add-in GPU Choice:

- NVIDIA RTX A2000 (12GB ECC VRAM): Enterprise driver stability, 70W max power.

- GeForce RTX 4060 LP (8GB VRAM): High consumer compute speed.

• Networking: 2x 10GbE SFP+ ports + 2x 2.5GbE ports (True enterprise hypervisor grade).
• RAM: Up to 64GB / 96GB DDR5
• Estimated Cost: $650 (Barebones) + $450 (RTX A2000 12GB) = $1,100 – $1,250
• Verdict: Best for enterprise network integration (dual 10G SFP+). RTX A2000 provides 12GB VRAM with ECC for 24/7 sovereign reliability.

Option D: NVIDIA Jetson AGX Orin Developer Kit (64GB)

• Architecture: NVIDIA Ampere (2048 CUDA cores, 64 Tensor cores)
• Memory: 64GB Unified LPDDR5 (2048-bit bus, 204.8 GB/s bandwidth)
• AI Performance: 275 TOPS (INT8)
• Power Envelope: 15W to 60W configurable
• Form Factor: Compact developer module + heatsink/carrier
• Estimated Cost: $1,999
• Verdict: Unlocks running large 30B–70B quantized models completely in unified GPU memory with virtually zero power draw, but runs ARM64 Linux (Ubuntu/L4T) rather than standard x86_64 Debian.

3. Recommended Capital Budget Plan

Tier                Model Recommendation                   Est. Budget     Primary Strength
---------------------------------------------------------------------------------------------------------
Recommended (x86)   Minisforum AtomMan G7 Ti (RTX 4070)    $1,450 - $1,600  Highest CPU/RAM (96GB), Plug-and-play
Workstation (10G)   Minisforum MS-01 + RTX A2000 (12GB)    $1,150 - $1,250  12GB ECC VRAM, Dual 10G SFP+
NUC Brand Direct    ASUS ROG NUC 970 (RTX 4070)            $1,850 - $2,050  Direct official Intel NUC lineage
Large Model Power   NVIDIA Jetson AGX Orin 64GB            $1,999           64GB Unified VRAM (Fits 70B models)

4. Immediate Next Step

Review preferred tier (Standard x86 RTX vs Dual-10G MS-01 vs Jetson Unified) to finalize funding allocation in REVENUE_RAMP_AND_TAX_MODEL.md.