Reducing Power BI model size: how VertiPaq works

Power BI uses the VertiPaq engine (also called xVelocity), an in-memory columnar engine. Understanding how it compresses data is the key to building small, fast models.

Compression by column

VertiPaq uses two main algorithms:

• Run-Length Encoding (RLE) — effective for columns with repeated values (e.g. Status with "Active"/"Inactive").
• Hash Encoding — builds a dictionary of unique values; each row stores only the dictionary index.

A column with 5 unique values across 10 million rows compresses far better than a column with 1 million unique values.

Spot problems with DAX Studio

DAX Studio is free and indispensable. Connect to the model and use Advanced → View Metrics:

-- See columns by size
SELECT TOP 20
    [TableName], [ColumnName],
    [TotalSize], [DataSize], [DictionarySize]
FROM $SYSTEM.DISCOVER_STORAGE_TABLE_COLUMNS
ORDER BY [TotalSize] DESC

Reduction actions

• Remove columns not used in reports or DAX measures.
• Reduce decimal precision (DECIMAL(18,2) vs DECIMAL(18,6)).
• Use correct data types: INT instead of STRING for surrogate keys.
• Disable auto date hierarchy in File → Options → Data Load.
• Use incremental refresh for large fact tables instead of importing everything.

Benchmark

A well-optimized model with 50 million rows can take less than 500 MB in memory. The same un-optimized model can take 5 GB. The difference is mostly in unnecessary columns and incorrect data types.

Conclusion

Knowing VertiPaq is not academic — it is practical. Smaller models refresh faster, cost less in Premium/Fabric, and users are happier with reports that open in seconds.