Reference
Glossary
Definitions of the acronyms, models, and technical terms used across the NAISH documentation.
Definitions of the acronyms, organisations, models, and technical terms used
throughout this documentation.
| Term | Definition |
|---|
| NAISH | Nigeria AI Scaling Hub – the national AI infrastructure initiative documented here |
| FMCIDE | Federal Ministry of Communications, Innovation & Digital Economy – client / partner |
| LBS | Lagos Business School (Pan-Atlantic University) – primary Gates grant recipient; manages and runs the Hub |
| DSN | Data Science Nigeria – technical implementation partner |
| GBB | Galaxy Backbone – physical hosting provider, Abuja |
| NCAIR | National Centre for AI and Robotics – technical sign-off authority per RFP |
| UduTech–Kasi Cloud Consortium | GPU hardware vendor (Ref: NAISH/INFRA/GPU-2026-02) |
| Dimagi | Organisation behind CommCare, the frontline data-collection platform |
| Term | Definition |
|---|
| H200 SXM5 | NVIDIA data-centre GPU; SXM5 socket form factor with 141 GB HBM3e |
| HBM3e | High-Bandwidth Memory (3rd-gen enhanced) – the on-package GPU memory |
| HGX | NVIDIA's multi-GPU baseboard platform; the 8 GPUs sit on one HGX baseboard |
| NVLink / NVSwitch | NVIDIA's high-bandwidth GPU-to-GPU interconnect fabric (900 GB/s full-mesh here) |
| SXM5 vs PCIe | SXM5 uses the HGX baseboard + NVSwitch (900 GB/s); PCIe is limited to ~64 GB/s over the host bus |
| TDP | Thermal Design Power – the power / heat a component is rated to dissipate |
| BTU/hr | British Thermal Units per hour – heat-rejection rate the cooling system must handle |
| CFM | Cubic Feet per Minute – airflow requirement for cooling |
| PDU | Power Distribution Unit – rack power feed (2× 15 kW metered here) |
| IPMI / Redfish | Out-of-band hardware management interfaces |
| Term | Definition |
|---|
| AGH | Africa GPU Hub – the vendor's optional orchestration layer |
| Kubernetes | Container orchestration platform |
| Volcano / Kueue | Batch scheduling and job-queueing systems for Kubernetes |
| Time-slicing | Splitting a single GPU across multiple workloads logically (via the NVIDIA Device Plugin) |
| vLLM | High-throughput LLM inference server with continuous batching |
| JupyterHub | Multi-user notebook environment |
| MLflow | Experiment and model-tracking system |
| MinIO | S3-compatible object storage, deployed on-node |
| Keycloak | Identity and access-management system |
| Prometheus / Grafana / DCGM Exporter | Metrics collection, dashboards, and GPU telemetry |
| Term | Definition |
|---|
| N-ATLaS-LLM 8B | Open-source Nigerian-language large language model (agricultural advisory) |
| Whisper-large-v3 | OpenAI's automatic speech-recognition (ASR) model (ORF assessment) |
| TrOCR / PaddleOCR | Optical Character Recognition models (health-record digitisation) |
| VRAM | Video RAM – GPU memory a model occupies while loaded |
| FP16 / INT8 | Numeric precisions for model weights; INT8 (quantised) roughly halves VRAM vs FP16 |
| LoRA / QLoRA | Parameter-efficient fine-tuning methods that reduce the compute needed to adapt LLMs |
| RLHF / DPO | Reinforcement Learning from Human Feedback / Direct Preference Optimisation – alignment methods |
| ASR | Automatic Speech Recognition |
| OCR / NLP | Optical Character Recognition / Natural Language Processing |
| Term | Definition |
|---|
| DHIS2 | District Health Information System 2 – standard health-data platform |
| EHR | Electronic Health Record |
| NDPA | Nigeria Data Protection Act |
| SOC 2 / ISO 27001 | Information-security compliance frameworks |
| AI POCUS | AI Point-of-Care Ultrasound |
| ORF | Oral Reading Fluency |
| Self-CAIRE | The digital health-coaching use case delivered via mDoc's "Kem" platform |
| ChatMRPT | Malaria intervention allocation copilot |
