Key improvements in the suite include enhanced performance for eigensolvers and singular value decomposition in RocSOLVER, and updates to the ROCm Offline Installer Creator that now supports Oracle Linux 8.10, 9.6, and SLES 15 SP7. The documentation has also been revised for clarity, including new tutorials aimed at AI developers and support for additional deep learning frameworks like the Deep Graph Library, Stanford Megatron-LM, and Verl.
Various updates have been made to ROCm's components, such as improved CPU Unit Occupancy tracking, GPU board voltage support, and enhanced math library functionalities across several modules. Significant changes include corrections to VRAM memory calculations and improvements in event recording and synchronization processes.
Moreover, ROCm 6.4.2 has introduced compatibility with the RDNA3 architecture-based Radeon RX 7700 XT GPU, and added support for SLES 15 SP7. However, it also marks the end of support for RHEL 9.5.
In the future, ROCm plans to transition the AMD SMI tool to a separate repository and phase out ROCm SMI, focusing on enhancing AMD SMI's features. Additionally, upcoming releases may see the deprecation of several profiling tools like ROCTracer and ROCProfiler in favor of the ROCprofiler-SDK.
The release notes provide detailed insights into component updates, resolved issues, and known problems, ensuring users are well-informed of the enhancements and any ongoing challenges.
Overall, ROCm 6.4.2 represents a significant step forward in AMD's commitment to optimizing its software for high-performance computing and AI applications, with a clear roadmap for future enhancements.
Further Developments
As ROCm continues to evolve, users can expect:1. Enhanced Framework Support: The addition of support for more deep learning frameworks, which may include emerging technologies and libraries that facilitate GPU acceleration.
2. Improved Documentation and Tutorials: Continued efforts to provide comprehensive resources for developers, particularly those new to AI and machine learning, ensuring a smoother onboarding process.
3. Performance Optimization: Regular updates aimed at refining the performance metrics and capabilities of existing tools, particularly in the realm of machine learning and computational tasks.
4. Community Engagement: Increased collaboration with the developer community for feedback on features and support for real-world applications, which can guide future developments.
5. Transition to New Architectures: As hardware evolves, ROCm will likely introduce support for new GPU architectures, ensuring compatibility and leveraging their capabilities for advanced computing tasks.
By keeping the community informed and engaged, AMD aims to ensure that ROCm remains a robust platform for developers and researchers working in high-performance computing and AI
ROCm 6.4.2 released
ROCm 6.4.2 has been released with several improvements and enhancements. The ROCm Compute Profiler now supports various data types and uses AMD SMI instead of ROCm SMI. It also adds 8-bit floating point metrics support for AMD Instinct MI300 series accelerators. RocSOLVER has improved the performance of eigensolvers and singular value decomposition. ROCm Offline Installer Creator 6.4.2 includes support for Oracle Linux 8.10 and 9.6, and SLES 15 SP7. ROCm documentation has been updated to provide clearer guidance for various user needs and use cases. The release also includes new tutorials for AI developers and supports additional deep learning frameworks such as Deep Graph Library, Stanford Megatron-LM, and Verl. ROCm 6.4.2 also adds support for SLES 15 SP7 and the RDNA3 architecture-based Radeon RX 7700 XT GPU.
ROCm 6.4.2 includes several changes to its components, including the addition of CPU Unit Occupancy information per process, GPU Board voltage support, new firmware PLDM_BUNDLE, improved math support for hipBLAS, hipBLASLt, hipFFT, hipfort, hipRAND, hipSOLVER, hipSPARSE, hipALUTION, hipBLAS, hipSOLVER, hipSPARSE, rocALUTION, hipBLAS, hipSOLVER, hipWMMA, tensile, primitives, tools for system management, HIP, HIPCC, ROCm Validation Suite, performance tests, ROCm Compute Profiler, ROCm Systems Profiler, ROCProfiler, ROCprofiler-SDK, ROCTracer, development tools, and runtimes. Key changes include corrected VRAM memory calculation, improved implementation of hipEventRecordWithFlags, improved implementation of hipEventSynchronize, and improved support for gfx1151 on Linux. Additionally, ROCm Compute Profiler now supports 8-bit floating point metrics for AMD Instinct MI300 GPUs, and ROCm Validation Suite now supports NPS2/DPX and NPS4/CPX partition modes for AMD Instinct MI300X.
