Viewing Multiple AI Engine Partitions - Viewing Multiple AI Engine Partitions - 2026.1 English - UG1702

Prerequisites: Building a Multi-Partition Design

A multi-partition design contains two or more independent AI Engine graphs. Compile each graph separately into its own partition of the AI Engine array. Use the enable-partition AI Engine compiler option to assign each graph to a contiguous range of columns and to give the partition a unique name:

[aie]
enable-partition=<START_COLUMN>:<NUM_OF_COLUMN>:<PARTITION_NAME>

For example, to compile a graph into a partition named pr0 that occupies columns 6 and 7, use:

[aie]
enable-partition=6:2:pr0

For a complete description of the option and the underlying partition concept, see Compiling AI Engine Graph for Independent Partitions in the AI Engine Tools and Flows User Guide (UG1076) .

Opening a Multi-Partition Design in the Analysis View

1. Open the Vitis Unified IDE, then select File > Open Summary, or click Open Summary in the Analysis view toolbar.
2. Select the top-level summary for the system: the xrt.run_summary that a hardware run produces, or the link_summary that v++ -l produces.
3. Vitis Analyzer reads the per-partition list from the summary, locates each partition's AI Engine compile-summary file, and loads every partition into a single Analysis view session.

You do not need to invoke Set Compile Directory for each partition. Selecting the run summary or link summary is sufficient, and every partition that the packager recorded becomes available immediately.

Visual Indicators for Partitions

When you load a multi-partition design, the Analysis view marks every partition consistently across reports:

• In the Graph view, a visible outline encloses each partition so you can identify which kernels and connections belong to the same partition.
• In tabular reports such as the Kernels table, a partition icon precedes each row, and the partition's namespace qualifies every object name. Two graphs that contain a kernel of the same name appear under different partition namespaces in the table, which removes ambiguity. The view qualifies net names the same way.

Aggregated Views Across Partitions

The following views aggregate data across every loaded partition.

The Analysis view supports cross-probing between the Graph view and the Trace view within a single partition (see Profiling the AI Engine in the AI Engine Tools and Flows User Guide (UG1076) Profiling the AI Engine), but it does not extend this cross-probing across partitions. To scope Graph-to-Trace cross-probing to one partition in a multi-partition design, open the AI Engine compile summary for that partition individually rather than the link summary. The link summary always produces the aggregated views described above.

For a description of the underlying compilation summary reports, see Viewing AI Engine Compilation Summary Reports.

Trace Configuration for Multi-Partition Designs

For hardware trace acquisition on a multi-partition design, use the following configuration:

• At most one partition can use PLIO trace offload, and that partition must be the first AI Engine partition in the design.
• Every other partition must use GMIO trace offload. Multiple partitions can use GMIO concurrently.

The Analysis view does not support a design in which two or more partitions each request PLIO offload.

Configure the trace offload mode on each partition through the AI Engine compiler trace options. For details on the trace options, see Event Trace Build Flow in the AI Engine Tools and Flows User Guide (UG1076). For details on how the Analysis view presents the resulting waveform, see Open Trace Summary Using Time Window.

Limitations

The following limitations apply to multi-partition designs in the Analysis view:

• Per-partition compile-summary reports do not aggregate. The reports listed under each partition's aiecompile_summary continue to use the per-partition data that the AI Engine compiler produces. Only the Graph, Array, and trace views aggregate across partitions.
• Performance metrics depend on load order. If you open the link summary first, the Analysis view might not report performance metrics. If you open an AI Engine compile summary first and add the link summary later, the Analysis view keeps the performance metrics for that partition visible but does not regenerate them for the other partitions. To see performance metrics for every partition, open the AI Engine compile summary for each partition before you add the link summary.
• Asymmetric loading rule between the link summary and the AI Engine compile summary. After you open a link summary in the Analysis view, you cannot add AI Engine compile summaries to that session because the link summary already populates them. The reverse works: you can open one or more AI Engine compile summaries and then add a link summary on top.
• Backward compatibility for designs built with earlier releases. Designs built with toolchains earlier than 2026.1 do not list their AI Engine compile summaries in the link summary, and the Analysis view cannot perform full multi-partition discovery on them. For these designs, open one AI Engine work directory at a time, as in earlier releases.