AMD uProf™ version 5.3 delivers significant improvements in stability, profiling accuracy, security hardening, and performance improvements across workflows. Key Enhancements include:
CPU Profiler
New IMS Metric : Added
IBS_[LD,ST]_L1_DTLB_REFILL_LATfor Zen4/Zen5 platforms. This metric helps to determine to what extent loads and stores are bottlenecked by TLB activity.Faster Translation : Optimized translation speed for modules with inline functions.
Custom Session Naming : Users can specify
Session NameusingAMDuProfCLIoption.Virtualized IBS Support : Enabled
vIBSsupport to enable IBS in virtualized environment for KVM hypervisor.Python profiling : Significant reductions in overhead that includes lower collection time overhead. This enhancement will be beneficial for long running and AI workloads
AMDuProfPCM
HTML Report Enhancements : Multiple Enhancements to HTML report.
PCIe Metrics Support : Added PCIe metrics for Zen3 server platforms.
Hotspots and Threading Analysis
Database Storage and Translation Optimization : Optimized database storage and translation overhead.
Thread‑Specific Wait-Time Collection : Added a new Linux CLI option,
--collect-thread, to enable collection of wait‑time data for a specific thread, improving precision in threading and latency analysis.
HPC
MPI Tracing Enhancements : Expanded Fortran name mangling coverage in MPI tracing agents.
OpenMP Analysis Fixes : Corrected OpenMP imbalance and synchronization calculations.
Improved MPI Reporting : Improved accuracy of MPI timing and rank level reporting.
New Inefficiency Metric : Introduced``Unused threads`` metric in inefficiency metrics to handle PRs with different number of threads
Faster Report Generation: Optimized report generation for OpenMP, MPI, and hybrid workloads.
Robust MPI Profiling : Prevented silent loss of MPI trace visibility in mixed language applications
AMDuProf GUI
Enhanced Timeline Visualization (Linux) : Enabled timeline visualization for function tracing sessions on Linux.
Improved Thread Observability : Added Thread Name and TID tooltip in “All Thread Timeline” with no truncation. This enables enhanced observability of execution flow at thread level
MPI Per‑Rank Analysis : Added per-rank analysis of MPI data.
Performance Optimizations : Optimized translation and report generation time.
Other Updates
AMDSystemCheck Utility (Linux) : Packaged AMDSystemCheck Utility On Linux. This utility provides underlying OS/BIOS/platform details including topology.
Reporting Performance Improvements: Reduced overall report generation time for large sessions.
Database Backend Upgrade : Integrated DuckDB as a high performance database backend, replacing SQLite as the default database. SQLite is still available for backward compatibility.