# Data Quality Issues in AI Supercomputers Dataset Based on the evidence gathered, here are the data quality issues identified: ## 1. **High Missing Data Rates** From step-3, significant null values exist across key measures: - **Hardware Cost**: 98 nulls (20% of 482 rows) - **Energy Efficiency (log)**: 76 nulls (16%) - **Power Capacity (MW)**: 60 nulls (12%) - **Chip quantity (primary)**: 54 nulls (11%) - **Max OP/s (log) and H100 equivalents**: 22 nulls each (5%) - **Chip quantity (secondary)**: 464 nulls (96% - essentially unused field) ## 2. **Precision Artifacts / Floating Point Errors** From step-2 and step-12, several values show floating-point precision artifacts: - `Chip quantity (primary)` max: **200000.00000000003** (should be 200,000) - `Anonymized Chinese System`: **100000.00000000001** chips - `Anonymized Chinese System`: **30000.000000000004** chips These appear to be calculation errors rather than true measurements. ## 3. **Inconsistent Date Formats** From step-11, the `First Operational Date Note` field contains mixed formats: - **YYYY-MM-DD** (precise dates): "2024-11-18" - **YYYY-MM** (month only): "2024-04", "2025-01" - **Quarter notation**: "Q4 2024", "Q4 2023" - **Empty/Other**: Some dates have no note This creates ambiguity about actual operational dates. ## 4. **Extreme Outliers** From step-15, 30 systems exceed the upper fence for chip quantity (>6,754.5 chips). Notable examples: | System | Chips | H100 Equiv | Power (MW) | Issue | |--------|-------|------------|------------|-------| | Anonymized Chinese System | 200,000 | — | 80.0 | Missing H100 equivalents despite having chips | | xAI Colossus Phase 3 | 200,000 | 275,796 | 352.352 | Extreme scale | | Anonymized Chinese System | 100,000 | 3,000 | — | Missing power data | | Anonymized Chinese System | 40,000 | 90 | 20.0 | **Suspiciously low H100 equivalents** | | Anonymized Chinese System | 40,000 | 500 | 30.0 | **Suspiciously low H100 equivalents** | ## 5. **Inconsistent Chip-to-Performance Ratios** From step-12, some systems show implausible efficiency: - **Anonymized Chinese System**: 7,000 chips → only **4.0 H100 equivalents** (0.06% efficiency) - **Anonymized Chinese System**: 40,000 chips → **90 H100 equivalents** (0.23% efficiency) Compare to typical systems where chip count and H100 equivalents are more proportional. ## 6. **Empty Primary Chip Type** From step-14, **199 systems** (41% of dataset) have **empty `Chip type (primary)`** field, yet still report chip quantities ranging from 200 to 200,000 chips. This is a critical missing identifier. ## 7. **Certainty Imbalance** From step-10: - **Confirmed**: 401 systems (83%) - **Likely**: 81 systems (17%) The "Likely" systems show much higher average values (avg H100: 9,554 vs 2,168), suggesting uncertain data for the largest systems. ## 8. **Future Dates** From step-2, the latest operational date is **2025-07-22** (future from typical data collection dates), though step-2 reports 0 "future_dates" by its criteria. This suggests planned rather than operational systems are included. ## 9. **Duplicate Name Issue** From step-3, there is **1 duplicate name** in the dataset, though the `Possible Duplicate` and `Possible Duplicate Of` fields are empty for all rows. ## Summary **Most Critical Issues:** 1. Missing chip type for 41% of systems 2. Floating-point precision errors in chip quantities 3. Implausible chip-to-performance ratios (some systems 1000x less efficient than expected) 4. High missing data rates (20% missing cost, 12% missing power) 5. Inconsistent date format documentation