🎓 Lesson 10
D5
Onboard GPU Selection for Underground Rigs
Choosing the right graphics processing unit (GPU) to install directly on mining rigs underground so AI models can process sensor data in real time without sending it to the surface.
🎯 Learning Objectives
- ✓ Analyze thermal derating curves to calculate maximum sustained GPU clock frequency at 52°C ambient
- ✓ Design a GPU-powered edge inference pipeline for real-time ore boundary classification using measured rig power budget and latency constraints
- ✓ Explain trade-offs between FP16 throughput, memory bandwidth, and ECC support in selecting GPUs for safety-critical grade control inference
- ✓ Apply IEC 60079-0 and MSHA 30 CFR Part 18 requirements to evaluate GPU enclosure compliance for gassy underground environments
📖 Why This Matters
Underground mines generate terabytes of real-time sensor data—from borehole LiDAR, hyperspectral core scanners, and microseismic arrays—but transmitting everything to the surface for AI analysis introduces unacceptable latency (>5 s) and bandwidth bottlenecks. Onboard GPUs let rigs classify ore/waste boundaries *while drilling*, enabling immediate cut-off grade adjustments and reducing dilution by up to 18% (as validated at BHP’s Jansen potash project). Selecting the wrong GPU risks thermal shutdown mid-shift, inference failures during critical stope mapping, or non-compliance with explosion-proofing mandates—making this not just an IT decision, but a mine planning and safety imperative.
📘 Core Principles
GPU selection hinges on four interdependent domains: (1) Computational: Measured in INT8/FP16 TOPS, required to meet inference throughput (e.g., ≥30 FPS for 1024×768 hyperspectral segmentation); (2) Thermal & Mechanical: Must operate continuously at 45–55°C ambient with no active airflow (conduction-cooled enclosures only); (3) Power: Drawn from DC 24–48 V vehicle bus; peak draw must stay ≤85% of available capacity after accounting for drill hydraulics, comms, and lighting loads; (4) Certification: GPUs deployed in gassy or dusty zones require either intrinsic safety (IS) design or encapsulation meeting IEC 60079-11 or flameproof housing per IEC 60079-1. Ruggedization (MIL-STD-810H shock/vibe, EN 50155 rail-grade EMI immunity) is non-negotiable—even if not in a classified zone—due to continuous jarring on development drifts.
📐 Thermal Derating Margin Calculation
GPU performance degrades as junction temperature rises above its thermal design point (TDP). Engineers must calculate the usable compute margin at site-specific ambient conditions using manufacturer-provided derating curves. This ensures inference latency remains within SLA even during prolonged operation in hot, stagnant air.
Effective Throughput Ratio (ETR)
ETR = 1 − [k × (T_amb − T_ref)]Calculates usable GPU computational throughput as a fraction of rated peak, adjusted for ambient temperature-induced thermal derating.
Variables:
| Symbol | Name | Unit | Description |
|---|---|---|---|
| ETR | Effective Throughput Ratio | dimensionless | Fraction of rated peak throughput achievable at operating ambient temperature |
| k | Derating Coefficient | °C⁻¹ | Manufacturer-specified throughput loss per degree Celsius above reference temperature |
| T_amb | Ambient Temperature | °C | Maximum expected air temperature surrounding GPU enclosure during operation |
| T_ref | Reference Temperature | °C | Temperature at which GPU peak throughput is rated (typically 25°C) |
Typical Ranges:
Conduction-cooled Jetson AGX Orin: 0.010 – 0.014 °C⁻¹
Embedded RTX A2000: 0.008 – 0.011 °C⁻¹
💡 Worked Example
Problem: NVIDIA Jetson AGX Orin (64GB) has FP16 peak = 204 TOPS at 25°C ambient. Datasheet shows linear derating: −1.2% TOPS per °C above 25°C. Rig operates at 52°C ambient in a sealed conduction-cooled enclosure. Calculate ETR and effective FP16 throughput.
1.
Step 1: Compute temperature delta = 52°C − 25°C = 27°C
2.
Step 2: Apply derating = 27°C × 1.2%/°C = 32.4% reduction
3.
Step 3: ETR = 1 − 0.324 = 0.676; Effective throughput = 204 TOPS × 0.676 = 137.9 TOPS
Answer:
The result is 137.9 FP16 TOPS, which exceeds the 110 TOPS minimum required for real-time stope boundary CNN inference (validated at Vale’s Sudbury Complex), confirming suitability.
🏗️ Real-World Application
At Rio Tinto’s Koodaideri Phase 2 iron ore operation, autonomous LHDs deploy onboard NVIDIA RTX A2000 (embedded variant) GPUs to run a U-Net model that segments real-time LiDAR + thermal fusion scans of muck piles. The GPU was selected after validating: (a) 105 W TDP fits within the LHD’s 220 W reserved edge budget; (b) Passive copper cold-plate cooling maintains <92°C GPU junction temp at 48°C ambient; (c) Enclosure certified to IEC 60079-15 (non-sparking protection) for Zone 2; and (d) Inference latency stable at 68 ± 5 ms over 12-hr shifts. This reduced ore loss by 12.3% in transition zones versus cloud-only inference.
📋 Case Connection
📋 Iron Ore Mine Sensor Fusion for Banded Iron Formation (BIF) Delineation
Conventional geophysics failed to resolve thin hematite bands (<2m) within jaspilite, causing 22% grade variance in ROM...