January 2026 (CodeLinaro/onnxruntime) delivered a focused set of power-management and efficiency improvements for the QNN Execution Provider. The work enables dynamic tuning of HTP power configurations and reduces unnecessary DSPQ polling, aligning resource usage with real workloads and preparing the ground for broader performance optimization.

December 2025: Delivered a critical resource-management improvement in ROCm/onnxruntime by aligning HTP power config ID lifecycle with session scope. Implemented per-session ID management, removed PerThreadContext, and introduced ManagedHtpPowerConfigId to guarantee a single power config ID per session. This change eliminates ID exhaustion under high-concurrency workloads, reduces per-thread overhead, and improves stability and scalability of multi-threaded inference. The work is captured in commit 8b81d9b13f94c1baeb2122a9405d33cf51968719 ([QNN-EP] - Tie HTP power config id lifetime to session (#26457)). Business value: more predictable resource usage, fewer runtime failures, and improved throughput under concurrent workloads.

November 2025 monthly summary for ROCm/onnxruntime focused on performance optimization of QNN Backend Manager power policy. Implemented Sustained High-Performance (SHP) power configuration by aligning SHP voltage votes with burst settings, eliminating the need for DSPQ polling and leveraging expected voltage mappings. The change was delivered via commit c30905d638418383b8d83b3b1bb65b7b42226f5a (PR #26465). The effort improves sustained performance predictability and efficiency in production workloads while simplifying power-management logic.

In October 2025, ROCm/onnxruntime delivered a major enhancement to the QNN Execution Provider by adding OpTrace profiling via the QNN System Profile API. This enables detailed performance debugging by producing a binary log compatible with qnn-profile-viewer, while preserving existing CSV profiling for backward compatibility. The work includes a new system profile serializer, API-versioning safeguards to support QNN API >= 2.29, and end-to-end tests that verify log generation and proper library loading. This feature accelerates performance diagnosis, reduces debugging time, and improves visibility into QNN EP workloads, delivering tangible business value for users relying on QNN-accelerated workloads.

September 2025: Focused on stabilizing QNN Execution Provider behavior in microsoft/onnxruntime. Delivered a targeted bug fix to RPC polling interval logic, aligning with performance modes and improving execution flow and resource management. The change reduces unnecessary polling and ensures predictable performance tuning when operating in burst mode.

Concise monthly summary for August 2025 focused on stabilizing the VTCM buffer sharing path in microsoft/onnxruntime through a targeted bug fix in the QNN-EP integration, delivering increased reliability and predictable memory behavior.

July 2025: Focused on QNN Execution Provider performance management enhancements in microsoft/onnxruntime. Delivered DSPQueue polling for burst-mode performance profiling, introduced new RPC power configurations with polling time, and added Efficient Mode API to set context priority based on workload type. These changes enhance performance tunability, readiness for burst workloads, and resource-aware scheduling in production deployments.

Monthly summary for 2025-06: Focused on delivering memory-optimized graph execution via VTCM backup buffer sharing in the QNN execution provider for OnnxRuntime. This work reduces memory overhead and improves resource utilization for large-scale models, enabling higher concurrency and more efficient production workloads. No explicit major bugs fixed in this period for microsoft/onnxruntime based on available data; feature-focused improvements support scalability and cost-efficiency.