## Scatter Plot: LongBench Accuracy vs KV Cache Size (Qasper Dataset)
### Overview
The chart visualizes the relationship between KV cache size (in GB) and LongBench accuracy for different key-value (KV) caching algorithms on the Qasper dataset. A Pareto Frontier line connects optimal data points, illustrating the trade-off between cache size and performance.
### Components/Axes
- **X-axis**: KV Cache Size (GB) ranging from 0.00 to 2.00 in 0.25 increments.
- **Y-axis**: LongBench Accuracy (16 to 23) in 1-point increments.
- **Legend**: Located at top-right, mapping colors/shapes to algorithms:
- Red squares: H₂O
- Blue circle: KIVI
- Green diamonds: SnapKV
- Yellow triangles: MiniKV
- Black line with x markers: Pareto Frontier
### Detailed Analysis
1. **H₂O (Red Squares)**:
- Points at (0.25, 17), (0.5, 17), (1.0, 20), (1.75, 20.5)
- Accuracy plateaus at 17 for 0.25–0.5 GB, then increases to 20.5 at 1.75 GB.
2. **KIVI (Blue Circle)**:
- Single point at (0.5, 22.5), lying on the Pareto Frontier.
3. **SnapKV (Green Diamonds)**:
- Points at (0.25, 16.5), (0.5, 19), (1.0, 22.5), (1.25, 23)
- Shows consistent improvement with cache size, reaching the Pareto Frontier at 1.0–1.25 GB.
4. **MiniKV (Yellow Triangles)**:
- Points at (0.1, 16.5), (0.25, 21), (0.5, 22), (0.75, 22.5)
- Rapid accuracy gains up to 0.75 GB, then plateaus.
5. **Pareto Frontier (Black Line)**:
- Connects optimal points: (0.1, 16.5) → (0.25, 21) → (0.5, 22.5) → (1.0, 22.5) → (1.25, 23)
- Represents the maximum achievable accuracy for each cache size.
### Key Observations
- **Optimal Performance**: KIVI (0.5 GB, 22.5) and SnapKV (1.25 GB, 23) lie on the Pareto Frontier, indicating peak efficiency.
- **Suboptimal Algorithms**: H₂O and MiniKV fall below the frontier, suggesting inefficiencies (e.g., H₂O’s 20.5 accuracy at 1.75 GB vs. SnapKV’s 23 at 1.25 GB).
- **Cache Size Trade-offs**: Larger caches (1.0–2.0 GB) generally improve accuracy but may not always align with the Pareto Frontier.
### Interpretation
The data demonstrates that **cache size and algorithm efficiency are interdependent**. While increasing cache size typically boosts accuracy, the Pareto Frontier highlights configurations where this trade-off is optimized. For example:
- **KIVI** achieves high accuracy (22.5) at a modest 0.5 GB, outperforming H₂O at 1.75 GB (20.5).
- **SnapKV** maximizes accuracy (23) at 1.25 GB, suggesting diminishing returns beyond this point.
- **H₂O** and **MiniKV** underperform relative to the frontier, indicating potential architectural or implementation limitations.
This analysis underscores the importance of selecting algorithms that align with the Pareto Frontier for resource-constrained environments, balancing cache size and computational efficiency.
