ROCm7 Logo Image © AMDROCm7 Logo (Image © AMD)

Transition to a Modular Architecture

The most significant new feature in Version 7.14.0 is the transition to a leaner Core SDK that includes only the essential runtime and development components. To support specific professional workflows, AMD has introduced optional, domain-specific extensions for artificial intelligence (AI), data science, and high-performance computing (HPC). This modular approach allows you to install only the components required for your specific project, reducing the software’s overall footprint and simplifying the installation process.

Expanded Hardware and Operating System Compatibility

The latest version extends support to multiple GPU architectures, operating systems, and virtualization environments:

GPU and APU Support

ROCm 7.14.0 now offers compatibility for various Ryzen AI processors, including:

  • Ryzen AI MAX+ PRO 495, MAX PRO 490, and MAX PRO 485 (gfx1151)
  • Ryzen AI 5 435, 5 430, 5 PRO 435 (gfx1153), and Ryzen AI 7 445 (gfx1153)

Operating System Updates

Support for RHEL 10.2 and RHEL 9.8 on Instinct and Radeon GPUs has been expanded. Additionally, Debian 13, as well as SUSE Linux Enterprise Server (SLES) 16 and 15 SP7, are now supported on the Instinct MI350P.

Virtualization and Partitioning

New virtualization configurations have been added for Radeon GPUs, including KVM passthrough for the Radeon AI PRO R9700S and SR-IOV for the Radeon PRO V710 on Ubuntu 24.04. For Instinct MI355X and MI350X deployments, the update optimizes multi-VF partition modes specifically for DPX and CPX compute partitions with NPS2 storage partitioning.

AI Frameworks and Inference

This release updates compatibility for several key AI frameworks to ensure stability and performance:

  • PyTorch: 2.12.0
  • TensorFlow: 2.21
  • JAX: 0.10.0
  • vLLM: 0.23.0
  • SGLang: 0.5.13

Developer Tools and Profiling Improvements

AMD has introduced several updates to the HIP API and the profiling suite to improve developer workflows.

Improvements to the HIP API

The update introduces Execution Context APIs that enable more efficient partitioning of GPU computing resources on a single device. To achieve better compatibility with CUDA, new APIs for asynchronous batch memory management and library management functions have been added. Additionally, HIP-Graph replay has been optimized to reduce overhead during asynchronous memory allocations.

Profiling and Telemetry

Starting with version 2.12, the ROCprofiler SDK now serves as the backend for the PyTorch profiler. Key new features include:

  • B Streaming Performance Monitors (SPM)B: Beta support for time-resolved hardware counter data on Instinct MI300X, MI325X, MI350X, and MI355X.
  • B Selective ProfilingB: You can now use ROCTx markers to capture GPU activity only in specific areas of your application.

The ROCm Systems Profiler now offers unified memory profiling to track page migrations and errors, as well as profiling of SDMA engine activity to identify data transfer bottlenecks in multi-GPU configurations.

System Management and Telemetry

The AMD SMI tool has been updated and now provides GPU metrics for temperature, clock speed, and utilization per partition. It now includes APU-specific CLI metrics and enables VRAM allocation and GTT optimization directly from the command line. For enterprise environments, the ROCm Data Center (RDC) tool adds 59 new telemetry fields covering power, PCIe activity, and status metrics, providing nearly the same feature set as the Device Metrics Exporter.

Library Updates and Deprecated Features

Several core libraries have been updated:

  • hipFile: A new library for Direct Storage I/O is included in the Core SDK and enables data transfers between memory and GPU memory, bypassing host-side copy operations on Instinct GPUs.
  • hipSPARSE: Now supports the Block Sparse Row (BSR) format in generic routines.
  • RCCL: Introduces a hierarchical AllGather algorithm for large, multi-node jobs and offloads collective data transfers to the GPU copy engine of MI355X GPUs.
  • rocSPARSE: Now provides support for incomplete LDLT factorization.

Finally, the ROCm Bandwidth Test (RBT) has reached the end of its life cycle and is no longer supported in this version; users are advised to switch to TransferBench or the ROCm Validation Suite.