v2.0.3

May 28th 2026

Added

• New Bandwidth Measurers:
  • Intel PCM: Added Intel PCM as a bandwidth measurement backend with CXL compatibility
  • Expanded LIKWID: Counter aggregation across multiple CPUs, automatic HBM sub-channel detection (SCHX), extra MBOX events, and full add-counters support
  • Intel VTune: Added as an alternative bandwidth measurement backend
• Instruction Sampling: New optional instruction sampling infrastructure for fine-grained execution analysis
• Adaptive Curve-Guided Pause Discovery: Automatically derives optimal pause timings from the bandwidth curve.

Improved

• Platform Support:
  • Expanded ISA Support: Added support for additional ISAs within existing architectures
  • ARM Systems: Bandwidth and latency measurements are now significantly more reliable
  • SubNUMA / HBM Latency: Improved latency measurements on SubNUMA and HBM systems
  • CXL Compatibility: Full support for CXL-attached memory via Intel PCM backend
• Measurement Stability:
  • Bandwidth Stabilization: Improved warmup handling with automatic bypass for low-throughput scenarios; more robust windowing and steady-state detection
  • Performance Counters: Extra counters now measured in the same command as bandwidth counters (reduced overhead); fixed counter value reporting for perf add-counters
• Topology & Binding:
  • Topology Detection: Enhanced NUMA and CPU topology discovery for more accurate system characterization
  • Memory Binding: Improved CPU and memory binding logic with better inheritance from numactl/taskset
• Assembly Generation:
  • New AddressingPolicy for flexible memory addressing modes
  • New StoreLinePolicy for configurable store line management
• Curve Processing: Smoother bandwidth/latency curve output with improved correction and visualization formatting
• CLI: Better argument validation, clearer error messages, and improved diagnostics

Fixed

• Various Minor bugs