Update: 2026-02-28 (05:36 AM)
Technical Intelligence Report: 2026-02-28
Executive Summary
- Linux Kernel Preparations for Zen 6: AMD has begun queuing patches for the Linux 7.1 kernel (expected April 2026) to support next-generation Zen 6 processors.
- Profiling Enhancements: The updates focus specifically on the
perfsubsystem, enhancing Instruction-Based Sampling (IBS) for future EPYC “Venice” CPUs, improving latency filtering and multi-socket profiling capabilities.
🤖 ROCm Updates & Software
[2026-02-28] AMD Prepares Linux For Instruction-Based Sampling Improvements With Zen 6
Source: Phoronix
Key takeaway relevant to AMD:
- Developer Impact: These updates are critical for developers and system administrators using EPYC “Venice” (Zen 6) systems, as they significantly improve the granularity and safety of low-level performance profiling via the Linux
perftool. - Security & Usability: The introduction of hardware-assisted privilege filtering allows unprivileged users to utilize IBS safely, removing previous software-based barriers.
Summary:
- AMD has submitted patches to the
tip/tip.gitbranch, targeting the Linux 7.1 kernel merge window in April 2026. - The patches introduce support for new Instruction-Based Sampling (IBS) capabilities found in upcoming Zen 6 processors.
- These improvements address race conditions, privilege filtering, and specific event sampling (latency/streaming stores).
Details:
- Target Architecture: The patches are designed for “future” AMD processors, explicitly identified as next-gen Zen 6 (likely EPYC Venice).
- Technical Improvements & Capabilities:
- Race Condition Elimination: Introduces an “Alternate disable bit” with control-only Model Specific Registers (MSRs). This eliminates the Read-Modify-Write (RMW) race condition present in existing
IBS_{FETCH|OP}_CTLMSRs. - RIP Bit 63 Filtering: Leveraging hardware-assisted privilege filtering via RIP (Instruction Pointer) bit 63 status. This enables IBS usage for unprivileged users without requiring complex software-based privilege filtering logic.
- Latency Thresholds: Adds a fetch latency threshold filter, allowing developers to capture only high-latency fetch events (filtering only when values exceed a programmable threshold).
- Streaming-Store Filter: Provides the ability to optionally record samples exclusively for instructions performing streaming stores.
- Remote Socket Indication: Adds support for indicating the data source for remote sockets during load/store instruction sampling, critical for optimizing NUMA-aware workloads on multi-socket EPYC servers.
- Race Condition Elimination: Introduces an “Alternate disable bit” with control-only Model Specific Registers (MSRs). This eliminates the Read-Modify-Write (RMW) race condition present in existing
- Timeline: Currently in the
perf/coreGit branch; expected to merge into the mainline kernel with Linux 7.1 in April.