ROCm Optiq (Beta) release history

ROCm Optiq (Beta) release history#

2026-03-27

6 min read time

Applies to Linux and Windows

Version

Release date

Beta 0.3.0

March 26, 2026

Beta 0.2.0

February 11, 2026

Beta 0.1.0

December 10, 2025

ROCm Optiq (Beta) 0.3.0#

Added#

ROCm Optiq for visualizing ROCm Compute Profiler’s data. New features include:

  • Summary View: Shows a high-level overview of the captured data.

    • Table: lists the top 10 longest-running kernels sorted by Total Execution Time.

    • Charts: Plot duration and invocation statistics across kernels.

    • Roofline Chart: plots kernel performance against empirical hardware ceilings to reveal the dominant performance bottleneck for all kernels.

  • Kernel Details: displays details of each kernel.

    • Kernel Selection Table: Lists kernels with GPU metrics. Use Add Metric to append additional GPU metric columns. Per-column search box accepts names or metric expressions (for example, metric > threshold). Click Apply Filters to execute; combine multiple filters to narrow the analysis.

    • Memory Chart: Shows memory transactions and throughput per cache hierarchy level for the selected kernel.

    • System Speed-of-Light: Displays key kernel-level performance metrics with unit, average, peak, and percentage of peak values.

    • Kernel Roofline Chart: Shows a kernel-specific roofline analysis to determine whether a kernel is compute-bound or memory-bound. Click the gear icon to access customization options.

  • Table View: provides a complete list of available metrics for the selected kernel.

  • Workload Details: Provides contextual information about the workload.

Changed#

Changes in ROCm Optiq for visualizing ROCm System Profiler traces:

  • System Topology tree was restructured to show hardware and software topologies.

  • Memory allocation activity tracks are now displayed in Timeline and System Topology Views.

  • RPD files populate Topology.

  • Multinode support: Time normalization for multi-node configurations.

Known issues#

Metrics that reference None return N/A#

If a metric expression contains None, ROCm Compute Profiler may ignore the metric value even when it isn’t None. As a result, ROCm-Optiq displays N/A for affected metrics.

  • System Speed of Light (0200)

    • VALU Active Threads

    • LDS Bank Conflicts/Access

    • vL1D Cache Hit Rate

    • L2 Cache Hit Rate

    • L2-Fabric Read Latency

    • L2-Fabric Write Latency

    • sL1D Cache Hit Rate

    • L1I Fetch Latency

  • Memory Chart (0300)

    • LDS Latency

    • VL1 Hit

    • VL1 Lat

    • VL1 Coalesce

    • VL1 Stall

    • sL1D Hit

    • sL1D Lat

    • IL1 Lat

    • L2 Rd Lat

    • L2 Wr Lat

  • Command Processor CPC/CPF (0500)

    • CPF Utilization

    • CPF Stall

    • CPF-L2 Utilization

    • CPF-L2 Stall

    • CPF-UTCL1 Stall

    • CPC SYNC FIFO Full Rate

    • CPC CANE Stall Rate

    • CPC ADC Utilization

    • CPC Utilization

    • CPC Stall Rate

    • CPC Packet Decoding Utilization

    • CPC-Workgroup Manager Utilization

    • CPC-L2 Utilization

    • CPC-UTCL1 Stall

    • CPC-UTCL2 Utilization

  • Workgroup Manager SPI (0600)

    • VGPR Writes

    • SGPR Writes

    • Not-scheduled Rate (Workgroup Manager)

    • Not-scheduled Rate (Scheduler-Pipe)

    • Scheduler-Pipe FIFO Full Rate

    • Scheduler-Pipe Stall Rate

    • Scratch Stall Rate

  • Compute Units Compute Pipeline (1100)

    • VALU Active Threads

    • MFMA Instruction Cycles

    • VMEM Latency

    • SMEM Latency

  • Local Data Share LDS (1200)

    • Bank Conflict Rate

    • LDS Latency

    • Bank Conflicts/Access

  • Scalar L1 Data Cache (1400)

    • Cache Hit Rate

  • Vector L1 Data Cache (1600)

    • Hit rate

    • Utilization

    • Coalescing

    • Stalled on L2 Data

    • Stalled on L2 Req

    • Stalled on Address

    • Stalled on Data

    • Stalled on Latency FIFO

    • Stalled on Request FIFO

    • Stalled on Read Return

    • Tag RAM Stall (Read)

    • Tag RAM Stall (Write)

    • Tag RAM Stall (Atomic)

    • Cache Hit Rate

    • Hit Ratio

  • L2 Cache (1700)

    • HBM Read Traffic

    • Remote Read Traffic

    • Uncached Read Traffic

    • HBM Write and Atomic Traffic

    • Remote Write and Atomic Traffic

    • Atomic Traffic

    • Uncached Write and Atomic Traffic

    • Read Latency

    • Write and Atomic Latency

    • Atomic Latency

    • Read Stall

    • Write Stall

    • Cache Hit

    • Read - PCIe Stall

    • Read - Infinity Fabric Stall

    • Read - HBM Stall

    • Write - PCIe Stall

    • Write - Infinity Fabric Stall

    • Write - HBM Stall

    • Write - Credit Starvation

workload_name is missing in sysinfo.csv when using --output-directory#

When you profile with the --output-directory option, the workload_name column in sysinfo.csv might be empty. This can prevent views in the ROCm Compute Profiler analysis database from joining tables based on workload_name, which makes system information unavailable.

ROCm Optiq (Beta) 0.2.0#

Added#

  • Summary View: Displays the top ten kernels by execution time using pie charts, bar charts, or tables.

  • Minimap: Provides a compact overview of event density and counter values across the entire trace, enabling rapid navigation of large datasets.

Changed#

  • Timeline View: Improved navigation and selection. Added context menu option to create a time range filter from a selected event or events.

  • Advanced Details Panel: Aggregate by Column drop-down groups the results by the selected column. Options to size columns to fit in Event Table and Sample Table. Event Details now shows the function call’s arguments, if available.

  • Time Range Filtering: Improved time range selection.

  • Histogram: Shows event density in two display modes: “Normalization: All Tracks” and “Normalization: Visible Tracks”.

  • Multi-node: Multi-node data and a new multi-database yaml file format are supported.

ROCm Optiq (Beta) 0.1.0#

Initial release of ROCm Optiq (Beta).

Added#

  • System Topology View: Displays a hierarchical representation of the hardware or system components, such as nodes, processes, as well as the GPU queues, memory operations, threads, and more that belong to them.

  • Timeline View: Shows CPU and GPU activities, events, and performance metrics in chronological order for a detailed temporal analysis. ROCm Optiq allows you to zoom, filter, and bookmark data for fine-grained inspection. You can correlate GPU workloads with in-application CPU events and performance with hardware resource usage, enabling easy identification and remediation of performance blockers.

  • Advanced Details Panel: Provides an in-depth view of profiling data, enabling you to analyze performance metrics and event-specific information. It offers SQL-like filters and group-by operations.

  • Histogram: Shows the event density across all visible tracks and highlights the zoomed-in region to quickly identify hotspots.

  • Time Range Filtering: Select a specific time interval to filter events and counter samples for focused analysis.

  • Event Search: Quickly locate target events.

Contents