LLM Efficiency Tool Diagrams

This document describes the diagrams needed for the article. Create these as PNG files in this folder.

1. architecture.png

Purpose: Show the three-stage measurement pipeline with distributed execution

Style: Flowchart with parallel branches, clean modern look

Content:

┌──────────────────────────────────────────────────────────────────┐
│                    MEASUREMENT PIPELINE                          │
└──────────────────────────────────────────────────────────────────┘
                              │
              ┌───────────────┴───────────────┐
              ▼                               ▼
┌─────────────────────────┐     ┌─────────────────────────┐
│     CONFIGURATION       │     │      EXECUTION          │
│  ───────────────────    │     │  ───────────────────    │
│  • Model selection      │────▶│  • HuggingFace Accelerate
│  • Hardware config      │     │  • Tensor parallelism   │
│  • Generation params    │     │  • Barrier sync         │
│  • YAML inheritance     │     └────────────┬────────────┘
└─────────────────────────┘                  │
                              ┌──────────────┼──────────────┐
                              ▼              ▼              ▼
                         ┌────────┐    ┌────────┐    ┌────────┐
                         │ GPU 0  │    │ GPU 1  │    │ GPU N  │
                         │ ────── │    │ ────── │    │ ────── │
                         │ Energy │    │ Energy │    │ Energy │
                         │ Tokens │    │ Tokens │    │ Tokens │
                         │ Memory │    │ Memory │    │ Memory │
                         └────┬───┘    └────┬───┘    └────┬───┘
                              │              │              │
                              └──────────────┼──────────────┘
                                             │
                                             ▼
                              ┌─────────────────────────┐
                              │      AGGREGATION        │
                              │  ───────────────────    │
                              │  • Late aggregation     │
                              │  • Raw results preserved│
                              │  • Flexible analysis    │
                              └─────────────────────────┘

Colours:

Config stage: Blue/teal
Execution stage: Green
Per-GPU boxes: Light grey with coloured borders
Aggregation: Purple/violet

Dimensions: ~800×600px

2. measurement-flow.png (optional)

Purpose: Show what happens during a single measurement run

Style: Vertical timeline / sequence diagram

Content:

┌─────────────────────────────────────────────┐
│ 1. INITIALISATION                           │
│    • Load model onto GPU(s)                 │
│    • Configure distributed backend          │
│    • Start CodeCarbon tracker               │
└─────────────────────────────────────────────┘
                      │
                      ▼
┌─────────────────────────────────────────────┐
│ 2. WARM-UP (discarded)                      │
│    • 3 dummy forward passes                 │
│    • Trigger lazy initialisations           │
│    • Stabilise GPU clocks                   │
└─────────────────────────────────────────────┘
                      │
                      ▼
┌─────────────────────────────────────────────┐
│ 3. MEASUREMENT                              │
│    ┌─────────────────────────────────────┐  │
│    │ for batch in dataloader:            │  │
│    │   outputs = model.generate(batch)   │  │
│    │   record_tokens(outputs)            │  │
│    └─────────────────────────────────────┘  │
│    • Energy tracked continuously            │
│    • Tokens counted per process             │
└─────────────────────────────────────────────┘
                      │
                      ▼
┌─────────────────────────────────────────────┐
│ 4. COLLECTION                               │
│    • Stop CodeCarbon                        │
│    • Gather per-GPU metrics                 │
│    • Save raw results to JSON               │
└─────────────────────────────────────────────┘

Colours: Gradient from top (light) to bottom (darker)

Dimensions: ~500×700px

3. energy-sources.png (optional)

Purpose: Pie chart or stacked bar showing energy breakdown

Style: Simple data visualisation

Content:

Typical Energy Breakdown (LLaMA-3B, 4×A100)

┌────────────────────────────────────────────┐
│  ██████████████████████████████████░░░░░░  │  GPU:  89%
│  ████░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░  │  CPU:   7%
│  ██░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░  │  RAM:   4%
└────────────────────────────────────────────┘

Or as a pie:

GPU (89%): Large slice, primary colour
CPU (7%): Small slice, secondary colour
RAM (4%): Tiny slice, tertiary colour

Dimensions: ~400×400px

Tools Suggestions

Excalidraw (excalidraw.com): Hand-drawn style, good for pipeline diagrams
Figma: Clean, professional look
diagrams.net (draw.io): Free, feature-rich
Mermaid: If you want code-defined diagrams

Export as PNG at 2x resolution for crisp display on retina screens.