## BarChart: Average Bandwidth Utilization
### Overview
The chart compares average bandwidth utilization across 10 different tasks or applications, measured as percentages. Three configurations are analyzed:
- **7x MIG 1g** (green bars)
- **7x MPS 13%** (light green bars)
- **7x timeslice** (blue bars)
A reference line (striped pattern) represents **1 task, full GPU** utilization.
### Components/Axes
- **X-axis**: Task/application names (e.g., `autodock_2vaa`, `qiskit_qv_30q`).
- **Y-axis**: Average bandwidth utilization (0% to 100%).
- **Legend**:
- Green: 7x MIG 1g
- Light green: 7x MPS 13%
- Blue: 7x timeslice
- Striped: Reference (1 task, full GPU)
### Detailed Analysis
1. **autodock_2vaa**:
- All bars (green, light green, blue) are **<1%**.
- Striped reference: **~1%**.
2. **autodock_3ef5**:
- All bars (green, light green, blue) are **<1%**.
- Striped reference: **~1%**.
3. **hotspot**:
- Green: **~5%**
- Light green: **~5%**
- Blue: **~1%**
- Striped reference: **~1%**.
4. **lamms_reaxff**:
- Green: **~25%**
- Light green: **~25%**
- Blue: **~15%**
- Striped reference: **~20%**.
5. **llama3**:
- Green: **~55%**
- Light green: **~50%**
- Blue: **~35%**
- Striped reference: **~35%**.
6. **llmc_shakespeare**:
- Green: **~25%**
- Light green: **~25%**
- Blue: **~15%**
- Striped reference: **~15%**.
7. **llmc_tinystories_bf16**:
- Green: **~50%**
- Light green: **~50%**
- Blue: **~30%**
- Striped reference: **~15%**.
8. **neks_turbpipeperiodic**:
- Green: **~30%**
- Light green: **~35%**
- Blue: **~25%**
- Striped reference: **~15%**.
9. **faiss_sift1m**:
- Green: **~10%**
- Light green: **~10%**
- Blue: **~5%**
- Striped reference: **~5%**.
10. **qiskit_qv_30q**:
- Green: **~75%**
- Light green: **~80%**
- Blue: **~75%**
- Striped reference: **~75%**.
### Key Observations
- **Highest utilization**: `qiskit_qv_30q` dominates with ~75–80% across all configurations, matching the reference (striped).
- **Moderate utilization**: `llama3` and `llmc_tinystories_bf16` show ~50–55% for green/light green, with blue at ~30–35%.
- **Low utilization**: `autodock_2vaa`, `autodock_3ef5`, and `hotspot` remain below 5%.
- **Striped reference**: Matches the 7x configurations in `qiskit_qv_30q` but is lower for other tasks.
### Interpretation
The data suggests that **qiskit_qv_30q** is the most resource-intensive task, utilizing nearly 80% of bandwidth across all configurations. This aligns with the reference (full GPU), indicating it may be a benchmark for maximum utilization. Other tasks like `llama3` and `llmc_tinystories_bf16` show significant utilization but remain below the reference. The **7x MIG 1g** and **7x MPS 13%** configurations generally outperform the **7x timeslice** in most tasks, except for `qiskit_qv_30q`, where all configurations converge. This could imply that the 7x MIG/MPS setups are more efficient for certain workloads, while the timeslice configuration lags behind except in the most demanding case.
