AMD GPU Driver (amdgpu) 31.60.0 release notes

AMD GPU Driver (amdgpu) 31.60.0 release notes#

The release notes provide release highlights and resolved issues since the previous AMD GPU Driver release (31.50.0).

Release highlights#

AMD GPU Driver 31.60 is an incremental update on the Linux 7.x amdgpu line. It is aimed at customers using the latest AMD Instinct™ compute platforms on supported Linux distributions.

For compatibility between AMD GPU Driver, ROCm, GPUs, and operating systems, see the Compatibility matrix.

Notable new features and improvements in AMD GPU Driver 31.60.0 include:

Compute and Instinct improvements#

  • Memory partitions (NPS) on gfx12.1: Exposes supported NPS modes and partition switching so multi-partition Instinct configurations can be selected and validated.

  • Dispatch-log profiling: Adds the kernel driver dispatch-log stream for gfx9.5 and gfx12, enabling queue-armed traces, firmware-notification polling, and reads in documented record formats.

  • AIS: When the AMD GPU driver initializes AIS, it registers the GPU’s VRAM with the P2PDMA kernel subsystem, generating MEMORY_DEVICE_PCI_P2PDMA metadata pages stored in host DRAM. A new amdgpu module parameter, ais_disabled, is added to allow non-AIS deployments to skip VRAM registration and reclaim the associated host memory.

    • ais_disabled=0 (default): AIS is initialized by default.

    • ais_disabled=1: AIS initialization is skipped.

    Check current state at /sys/module/amdgpu/parameters/ais_disabled. For more information, see AIS (AMD Infinity Storage).

Broader hardware enablement#

  • Harvested IP discovery: Reads LSDMA/MMHUB harvest and reserved-memory tables so partially populated or partitioned parts enumerate correctly.

  • SOC 1.0 reset: Adds a platform reset handler and revision identification from IP discovery.

Reliability, Availability, and Serviceability (RAS)#

  • ECC-aware compute reset: Informs the kernel driver when a reset was caused by an ECC event, and extends poison-consumption handling for A+A SR-IOV guests.

  • Bad-page and EEPROM accounting: Improves bad-page counts/thresholds and records data-source timestamps in EEPROM.

Resolved issues#

The following issues have been resolved in this release:

  • Resolved hangs and races when destroying or restoring a user-mode queue, including a restore that could run before all GPU virtual addresses were mapped.

  • Resolved hangs and races on queue destroy, restored all mappings before replay, validated ring pointers, and added timeline synchronization-object signaling after preemption or reset.

  • Resolved the kernel driver shared-virtual-memory migration for VM ranges with holes, including a leak on the copy-to-RAM error path.

  • Resolved CPER error-record retrieval and count reporting, and driver unload failure when UniRAS is enabled.

Contents