Enterprise guide to semantic model design, calculation groups, field parameters, composite models, Direct Lake, governance, and shared model architecture.
Power BI Semantic Model Enterprise Guide 2026 — enterprise reference guide from EPC Group, built since 1997 of Microsoft consulting engagements at Fortune 500 scale. Covers architecture, governance, compliance, pricing benchmarks, and implementation timelines for the Microsoft ecosystem.
Last updated by Errin O'Connor, Founder & Chief AI Architect, EPC Group
What is a semantic model in Power BI? A Power BI semantic model is a centralized data layer. It defines various elements for reporting, including:
In 2023, Microsoft renamed it from "dataset" to emphasize its role as a meaning-based layer between raw data and business insights.
The semantic model is the single source of truth for your data. A well-designed model can support numerous reports, from dozens to hundreds. It offers:
The semantic model is crucial for your Power BI deployment. Every report, dashboard, and metric relies on the quality of this model. A poorly designed model can lead to:
In contrast, a well-designed model delivers fast and reliable enterprise-grade business intelligence.
EPC Group has designed and optimized Power BI semantic models for Fortune 500 organizations across healthcare, financial services, and government. This guide covers every aspect of enterprise semantic model design — from the fundamental rename from "dataset" to advanced features like calculation groups, field parameters, composite models, and Direct Lake in Microsoft Fabric.
Are you designing a new semantic model, migrating from old datasets, or optimizing an existing model for Fabric? This guide outlines the enterprise methodology that EPC Group uses for every engagement.
Microsoft's terminology has evolved, and the distinctions matter for architecture decisions. Understanding what each component does — and does not do — prevents costly design mistakes.
| Component | What It Is | Engine | Best For |
|---|---|---|---|
| Semantic Model (Dataset) | In-memory analytical model with tables, relationships, DAX measures, RLS, and calculation groups | Vertipaq (columnar compression) | Primary analytics layer — reports, dashboards, metrics |
| Datamart | Self-service relational database with SQL endpoint, no-code ETL, and auto-generated semantic model | Azure SQL (relational) | Teams needing SQL access to curated data subsets |
| Lakehouse (Fabric) | Delta Lake storage with SQL analytics endpoint and auto-created semantic model | Spark + SQL + Direct Lake | Enterprise data platform with unified storage in OneLake |
| Warehouse (Fabric) | Full T-SQL data warehouse with cross-database queries and semantic model | Distributed SQL engine | Complex transformations requiring full T-SQL support |
Key Takeaway: Every Fabric item, including Lakehouse, Warehouse, and Datamart, automatically generates a default semantic model. However, these default models often lack optimized relationships, measures, and governance.
EPC Group develops a custom semantic model on top of Fabric storage. We view the auto-generated model as a starting point rather than a final product.
These foundational principles guide every semantic model EPC Group builds. Violating any one creates performance, governance, or trust problems.
Every semantic model starts with a star schema — fact tables with numeric measures surrounded by dimension tables with descriptive attributes. The Vertipaq engine is optimized specifically for this pattern.
One certified semantic model per business domain. All reports connect to the shared model via live connection. No duplicate models, no conflicting numbers.
Replace hundreds of duplicate measures with reusable calculation items. Time intelligence, currency conversion, and scenario analysis as modular, maintainable groups.
Choose Import, DirectQuery, composite, or Direct Lake based on data volume, freshness requirements, and capacity constraints. No one-size-fits-all.
Endorsement levels (Promoted, Certified), workspace permissions, lineage tracking, and a formal certification process with documented quality criteria.
Publish certified models to dedicated workspaces. Report authors connect via live connection. Centralized security, single refresh schedule, consistent metrics.
Relationships impact the functionality of filters in your semantic model. A misconfigured relationship can lead to inaccurate numbers and poor performance. To ensure accuracy, EPC Group examines each relationship during our semantic model audit.
Single-direction cross-filtering (dimension filters fact) is the recommended default setting. Enable bidirectional filtering only when a specific visual requirement calls for it.
In those cases, use CROSSFILTER() in DAX instead of model-level bidirectional settings.
Bidirectional relationships can lead to:
When a single dimension table links to a fact table through multiple columns, use one active relationship. For the other connections, use USERELATIONSHIP() in DAX.
For instance, a Date dimension can connect to:
Do not duplicate the dimension table. Duplicates waste memory and create maintenance issues. Instead, use a single Date dimension with three relationships:
Calculation groups improve the maintainability of enterprise semantic models. They remove the need for many duplicate measures. Instead, a single reusable group of calculation items can dynamically change any measure.
A calculation group is a unique table with one column for calculation items. It applies DAX transformations to any measure chosen in a visual.
For example:
The measure itself remains unchanged; the calculation group wraps it dynamically.
Items: YTD, QTD, MTD, PY, PY YTD, YOY, YOY%, Rolling 12 Months
Replaces 8 variants per measure — 30 base measures x 8 = 240 measures reduced to 30 + 1 group
Items: USD, EUR, GBP, JPY, Local Currency
Replaces 5 variants per measure — eliminates currency-specific measures entirely
Items: Actual, Budget, Forecast, Variance, Variance %
Replaces 5 variants per measure — enables dynamic actual-vs-budget on any KPI
Items: Per Unit, Per 1000, Percentage of Total, Running Total, Moving Average
Replaces 5 variants per measure — report authors apply formatting without new measures
Enterprise Impact: EPC Group implemented calculation groups for a Fortune 500 financial services client with over 1,200 measures. After consolidation, the client had:
Model maintenance time decreased from 40 hours per month to just 8 hours. Additionally, measure consistency issues, such as differing YTD formulas across departments, were completely resolved.
Field parameters let report users modify the measures or columns displayed in visuals. This can be done without editing the report itself.
This feature transforms static, single-purpose pages into dynamic, multi-purpose dashboards.
As a result, organizations can:
Limitation: Field parameters create a disconnected table with NAMEOF(). They do not support dynamic security, meaning RLS does not apply to field parameter selections.
Test thoroughly before deploying in compliance-sensitive environments. EPC Group validates field parameter behavior against RLS in every deployment.
Choosing the right storage mode is crucial in designing a semantic model. This decision significantly impacts performance and usability. The best option depends on several factors:
| Storage Mode | Performance | Data Freshness | Best For |
|---|---|---|---|
| Import | Fastest (in-memory Vertipaq) | Scheduled refresh (minutes to hours) | Datasets under 1GB, dashboards needing maximum speed |
| DirectQuery | Depends on source (slower) | Real-time (every query hits source) | Real-time requirements, source handles query load |
| Composite | Fast dimensions + live facts | Hybrid (cached dimensions, live facts) | Datasets 1-100GB, balance speed and freshness |
| Direct Lake (Fabric) | Near-Import speed | Near-real-time (reads OneLake Parquet) | Fabric environments, eliminating refresh pipelines |
Direct Lake solves the main challenge in enterprise Power BI: data refresh. It eliminates the need for importing data into the Vertipaq engine. This traditional method requires:
In contrast, Direct Lake reads Delta/Parquet files directly from OneLake. The data is already available, allowing the semantic model to read it seamlessly.
EPC Group worked with a client who managed a 50GB dataset. This dataset needed a 45-minute import refresh twice a day. We successfully migrated the semantic model to Direct Lake. The results were impressive:
Direct Lake runs on any Fabric F SKU; guardrails for Parquet files, row groups, rows, model size and memory scale with the SKU, and F64 lifts the model-size cap. It also requires data to be stored in Delta format.
EPC Group includes the Lakehouse-to-Direct Lake pipeline as a standard feature in every Fabric deployment.
Performance optimization is not an afterthought — it is a design discipline. Every decision from schema to DAX affects query speed, refresh time, and capacity cost.
Every column in the semantic model consumes memory — even columns referenced by zero visuals. Audit with DAX Studio VertiPaq Analyzer and remove all unused columns.
Impact: 20-40% memory reduction
Replace text keys with integer keys for all relationships. Integer comparisons are 3-5x faster than text comparisons and compress more efficiently in Vertipaq.
Impact: 30-50% faster joins
For datasets over 1GB, configure incremental refresh to process only new and changed partitions. Requires a DateTime column and query folding support.
Impact: 80-98% refresh time reduction
Use CALCULATE with explicit filters, avoid FILTER() on large tables, prefer SUMMARIZECOLUMNS, minimize iterator functions on high-cardinality columns.
Impact: 40-70% query speed improvement
Create pre-aggregated summary tables for executive dashboards. Power BI automatically routes queries to the smallest table that satisfies the visual.
Impact: 70-90% dashboard load improvement
Ensure Power Query transformations fold to the source database. Unfolded steps require Power BI to download raw data before transforming — exponentially slower.
Impact: 50-80% refresh improvement
Without proper governance, semantic model sprawl can happen. Organizations might create several competing models that yield different results for the same metrics. To address this issue, EPC Group uses a three-tier governance framework for each enterprise deployment.
Exploratory
Personal workspace models for ad-hoc analysis. Not shared, not governed, not trusted for decisions. Any user can create.
Promoted
Team-level models that have been reviewed by the data team. Published to shared workspaces. Useful but not yet certified against organizational standards.
Certified
Enterprise-grade models that meet all quality, accuracy, security, and documentation standards. Only designated certifiers (Power BI admins) can certify. The gold standard.
The shared semantic model pattern is a key governance decision for enterprise Power BI. It stops each report author from making their own data model. This prevents the creation of over 50 models that may have conflicting numbers.
Instead, a centralized team is responsible for building and certifying shared models. All report authors connect to these certified models, ensuring consistency and accuracy across reports.
Enterprise Impact: EPC Group implemented shared semantic models for a healthcare system with 12 hospitals. Previously, there were:
After the implementation, the results included:
The report count remained over 200, but the model count decreased by 95%. Additionally, the monthly refresh capacity cost dropped by $4,200.
Our 4-phase approach takes any organization from ungoverned model sprawl to a certified, optimized, shared semantic model architecture.
Week 1
Inventory all existing datasets, their refresh schedules, RLS configurations, measure definitions, and downstream reports. Identify duplicate models, conflicting measures, and performance bottlenecks using DAX Studio and ALM Toolkit.
Deliverable: Semantic model landscape audit with consolidation roadmap
Week 2
Design the target star schema with fact/dimension tables, relationships, calculation groups, field parameters, and storage modes. Define the shared model workspace architecture, endorsement criteria, and governance processes.
Deliverable: Semantic model design document with governance framework
Weeks 3-5
Build the certified semantic models in Tabular Editor. Implement calculation groups, RLS, incremental refresh, and composite/Direct Lake storage. Migrate existing reports to connect to shared models.
Deliverable: Certified semantic models with migrated reports
Week 6
Performance test every connected report. Validate data accuracy against source systems. Certify models through the endorsement framework. Train model owners on maintenance and governance workflows.
Deliverable: Certified production models with trained governance team
A semantic model in Power BI is the unified data layer that defines tables, relationships, measures, hierarchies, and business logic for reporting. Previously called a "dataset," Microsoft renamed it to "semantic model" in late 2023 to better reflect its purpose — it provides a semantic (meaning-based) layer between raw data sources and report visuals. The semantic model contains the star schema design, DAX measures, row-level security rules, calculation groups, and field parameters that transform raw data into business-ready analytics. EPC Group designs semantic models as the single source of truth for enterprise analytics — one well-governed model that serves dozens or hundreds of reports.
A semantic model and dataset are the same thing — Microsoft renamed "dataset" to "semantic model" in 2023. Both refer to the data model containing tables, relationships, and DAX logic published to the Power BI Service. A datamart is a self-service, fully managed relational database within Power BI Premium that allows analysts to create SQL-queryable data stores without Azure SQL or Synapse. Key differences: semantic models use Vertipaq in-memory compression and DAX; datamarts use a SQL endpoint with T-SQL queries. Semantic models are the primary analytics layer for reports. Datamarts are useful when teams need SQL access to a subset of enterprise data. EPC Group recommends semantic models as the primary analytics layer and datamarts only for teams that require direct SQL access.
Calculation groups are reusable sets of DAX calculation items that modify how measures behave — eliminating the need to create dozens of duplicate measures for common transformations like time intelligence. For example, instead of creating "Sales YTD," "Profit YTD," "Revenue YTD," and "Cost YTD" separately, you create one calculation group called "Time Intelligence" with items like YTD, QTD, MTD, PY, PY YTD, and YOY%. Every measure in the model can then be combined with any calculation item. This reduces a 200-measure model to 30 base measures plus a single calculation group. EPC Group implements calculation groups in every enterprise semantic model — they reduce measure count by 60-80% and dramatically simplify maintenance.
Field parameters allow report consumers to dynamically switch which columns or measures appear in visuals without editing the report. Users select from a dropdown (e.g., "Revenue," "Profit," "Units Sold") and the visual updates to show the selected metric. Under the hood, field parameters create a disconnected table with NAMEOF() references to measures or columns. Use cases: executive dashboards where users toggle between KPIs, comparison visuals where users pick which dimensions to slice by, and mobile reports where screen space is limited. EPC Group uses field parameters in 90%+ of executive dashboards — they reduce the number of report pages by 40-60% while giving users more flexibility.
Direct Lake is a new storage mode in Microsoft Fabric that reads Parquet files directly from OneLake into the Vertipaq engine — combining the performance of Import mode with the freshness of DirectQuery. Unlike Import (which copies data during refresh) or DirectQuery (which queries the source live), Direct Lake reads columnar Parquet files on demand without a data copy step. Benefits: near-instant "refresh" because data is already in OneLake, query performance close to Import mode, no separate ETL pipeline to maintain. Limitations: requires Fabric capacity (F64 or higher for production), data must be in Delta/Parquet format in OneLake, and not all DAX functions are supported in the initial framing. EPC Group is migrating enterprise clients from Import and DirectQuery to Direct Lake as part of Fabric adoption — typical refresh times drop from 30-60 minutes to under 30 seconds.
Power BI provides two levels of semantic model endorsement: Promoted (any dataset owner can promote their model as recommended) and Certified (only designated certifiers approved by the Power BI admin can certify a model as meeting organizational quality standards). Governance best practices: 1) Establish clear certification criteria — data accuracy validation, performance benchmarks, RLS implementation, documentation requirements, 2) Create a semantic model registry that tracks all certified models, their owners, refresh schedules, and downstream reports, 3) Use workspace-level permissions to control who can build reports against certified models, 4) Enable lineage view to track data flow from source through semantic model to reports. EPC Group establishes governance frameworks that typically include a 5-step certification process with automated validation checks.
Shared semantic models (also called shared datasets) allow multiple reports across different workspaces to connect to a single published semantic model — creating a true single source of truth. Without shared models, every report author creates their own data model, leading to conflicting numbers, duplicated refresh schedules, and ungovernable sprawl. With shared models, one team maintains the certified semantic model, and report authors use "Live connection" or "DirectQuery to Power BI dataset" to build reports against it. Benefits: consistent metrics across the organization, reduced capacity consumption (one model instead of 50 copies), centralized security through model-level RLS, and simplified governance. EPC Group implements shared semantic model architectures for every enterprise client — a typical deployment has 5-10 certified models serving 200+ reports.
Composite models allow a single semantic model to combine multiple storage modes: Import (fast, cached), DirectQuery (real-time, source-queried), and Dual (both). The enterprise strategy: Import small dimension tables (products, customers, dates — fast filtering), keep large fact tables in DirectQuery (transactions, events — always current, no refresh needed), and use aggregation tables in Import mode for high-level visuals while DirectQuery serves detail-level drill-through. Composite models also support DirectQuery to Power BI datasets — enabling you to extend a certified shared semantic model with additional local tables without duplicating the entire model. EPC Group designs composite model architectures for datasets between 1-100GB, balancing performance, freshness, and capacity consumption.
Enterprise semantic model performance optimization covers five areas: 1) Schema design — star schema with integer keys, minimal columns, proper cardinality, 2) DAX efficiency — use CALCULATE with explicit filters, avoid FILTER() on large tables, prefer SUMMARIZECOLUMNS over ADDCOLUMNS/SUMMARIZE, 3) Storage modes — use composite models with Import dimensions and DirectQuery facts, or migrate to Direct Lake in Fabric, 4) Refresh optimization — implement incremental refresh with partition management, use query folding to push transformations to source, 5) Capacity management — right-size Premium/Fabric capacity, monitor with Capacity Metrics app, set max memory per dataset limits. EPC Group performance audits typically improve semantic model query times by 50-80% and reduce refresh duration by 60-90%.
Enterprise Power BI implementation, optimization, and managed services from EPC Group.
Read moreDeep technical guide to star schema design, relationships, calculated columns vs measures, and performance optimization.
Read moreFull-stack Microsoft analytics: Fabric, Power BI, Azure AI, and enterprise operating models.
Read moreSchedule a free semantic model assessment with EPC Group. We will review your current models and find ways to consolidate them. Our goal is to provide a shared semantic model architecture that:
A Power BI semantic model (formerly dataset) is the governed data layer. It defines measures, relationships, hierarchies, and security.
This guide includes:
Last updated: June 2025 · Read time: 22 min
Enterprise guide to semantic model design, calculation groups, field parameters, composite models, Direct Lake, governance, and shared model architecture.
The semantic model is essential for your Power BI deployment. It impacts every report, dashboard, and metric. A high-quality semantic model is vital for achieving overall success.
A poorly designed model produces slow reports, inconsistent numbers, and ungovernable analytics sprawl. A well-designed model produces fast, trustworthy, enterprise-grade business intelligence.
Relationships affect how filters operate within your semantic model. If a relationship is misconfigured, it can result in incorrect numbers or poor performance, or both.
EPC Group carefully examines each relationship in our semantic model audit.
Calculation groups improve the maintainability of enterprise semantic models. They remove the need for many duplicate measures. Instead, a single reusable group of calculation items can dynamically change any measure.
EPC Group worked with a client who managed a 50GB dataset. This dataset required a 45-minute import refresh twice a day. We migrated the semantic model to Direct Lake, which led to significant improvements:
Direct Lake needs a Fabric capacity of F64 or higher for production workloads. It also requires data to be stored in Delta format.
EPC Group includes the Lakehouse-to-Direct Lake pipeline as a standard part of every Fabric deployment.
The shared semantic model pattern is crucial for enterprise Power BI governance. It stops each report author from building their own data model. Without this, organizations can end up with over 50 models that show conflicting numbers.
Instead, a centralized team is responsible for building and certifying shared models. All report authors connect to these certified models, ensuring consistency and accuracy across reports.
Our 4-phase approach takes any organization from ungoverned model sprawl to a certified, optimized, shared semantic model architecture.
A semantic model (formerly dataset) serves as the governed data layer for Power BI reports. It defines important elements such as:
All reports linked to the model share one source of truth.
DirectLake mode allows you to read Parquet files stored in OneLake. You can access these files without importing data into Vertipaq. This mode offers query speeds that are almost as fast as Import mode. Additionally, it eliminates the costs associated with refresh windows.
To use DirectLake mode, you need:
Use descriptive field names, add descriptions to every measure and dimension, configure synonyms for common terms, and remove unused columns. These steps directly improve Copilot answer accuracy.
Import loads data into Vertipaq. This allows for quick queries, but the data is only as fresh as the last refresh. DirectQuery sends live queries to the source, ensuring the data is always current, but it is slower. DirectLake is the third option that is native to Fabric.
Yes, this applies to any model that includes sensitive, role-specific, or regulated data. Row-Level Security (RLS) is defined in the semantic model. It impacts all reports and Copilot responses associated with that model.
EPC Group has completed over 1,500 Power BI deployments for Fortune 500 companies and clients in regulated industries. Our architects literally wrote the book on Power BI.
Founder Errin O'Connor is a first awarded in 2003. He was also a member of the original Power BI Beta Team.
AI assistant — not human