ROCm 7.0.0 is a major release that brings significant updates and new features across its comprehensive platform, focusing on expanded hardware support, enhanced deep learning framework integration, improved compiler capabilities, and refined tooling. The release introduces support for AMD Instinct MI355X and MI350X GPUs, expands OS compatibility to include Ubuntu 24.04.3 and Rocky Linux 9, and enhances support for popular deep learning frameworks such as PyTorch and TensorFlow. Additionally, ROCm 7.0.0 features improved compiler capabilities, including the introduction of the AMD Next-Gen Fortran compiler and enhanced clang/llvm support. With these improvements, ROCm 7.0.0 offers a strong and adaptable platform for AI, machine learning, and high-performance computing tasks, speeding up the creation of advanced applications.
ROCm 7.0.0 released
The latest release of the Radeon Open Compute (ROCm) ecosystem, ROCm 7.0.0, brings significant updates and new features across its comprehensive platform. This major release focuses on expanded hardware support, enhanced deep learning framework integration, improved compiler capabilities, and refined tooling to further advance AI, machine learning, and high-performance computing workloads.
Key Highlights of ROCm 7.0.0:
Hardware and OS Support
- Expanded GPU Support: ROCm 7.0.0 introduces support for AMD Instinct MI355X and MI350X GPUs, providing enhanced performance capabilities.
- Enhanced OS Compatibility: The release expands OS support to include Ubuntu 24.04.3 and Rocky Linux 9, ensuring broader compatibility with existing infrastructure.
- KVM Passthrough and VMware ESXi Support: KVM Passthrough is now available for MI350X and MI355X GPUs, while VMware ESXi 8 support has been introduced for MI300X GPUs, enabling seamless integration into virtualized environments.
- Retirement of Legacy OS Versions: Ubuntu 24.04.2 and SLES 15 SP6 are no longer supported in ROCm 7.0.0 to focus on modern operating systems.
Deep Learning and AI Frameworks
- Updated Framework Support: PyTorch 2.7, JAX 0.6.0, Megatron-LM, TensorFlow 2.19.1, ONNX Runtime 1.22.0, and Triton 3.3.0 have been updated to ensure compatibility with the latest frameworks.
- Introduction of vLLM Support: ROCm 7.0.0 introduces support for vLLM (vectors of Long Latency Memory) with OCP FP8 data type and FP4 precision for Llama 3.1 405B, enhancing performance in large-scale AI workloads.
- Official Support for Ray and llama.cpp: Official support has been added for these frameworks to provide a seamless user experience.
Component Separations and Consolidations
- Separate Distribution of AMD GPU Driver: The AMD GPU Driver (amdgpu) is now distributed separately from the ROCm stack, allowing for more flexible integration with other ecosystems.
- Consolidation of ROCm Library Repositories: The ROCm library repositories are being consolidated into a single repository under rocm-libraries to simplify maintenance and updates.
HIP API and Runtime Enhancements
- Improved Cross-Platform Programming Compatibility: The HIP API has been improved for cross-platform compatibility, requiring recompilation of existing applications.
- OCP MX Floating-Point Support: OCP MX floating-point data types (FP4, FP6, FP8) are now supported in ROCm 7.0.0.
- Enhanced Runtime Logging and constexpr Operators: HIP runtime logging has been enhanced, and constexpr operators for FP16/BF16 have been introduced.
- New Device Capability Attribute: A new device capability attribute is available to report the number of compute dies (chiplets, xcc) in MI350X/MI355X GPUs.
Compiler and Tooling Updates
- Introduction of AMD Next-Gen Fortran Compiler: The AMD Next-Gen Fortran compiler (llvm-flang) has been introduced in ROCm 7.0.0.
- Enhanced clang/llvm Support: Enhanced clang/llvm support is available, including separate debug file generation for device code and improved llvm-strip for AMD GPU device code objects.
- Comgr Virtual File System (VFS): Comgr now supports an in-memory virtual file system (VFS) for intermediate compilation steps.
- Parallel Code Generation: Parallel code generation is now default for HIP with full LTO (-fgpu-rdc).
Library and Data Type Updates
- Microscaling (MX) Support: Support has been added for Microscaling (MX) data types FP4, FP6, and FP8 in various libraries like Composable Kernel and hipBLASLt.
- OCP FP8 Format Support: Several libraries have been updated to support the Open Compute Project (OCP) FP8 format.
- MIGraphX Improvements: MIGraphX has received improvements, including OCP FP8 support, PyTorch 2.7 integration, and improved performance for Generative AI models.
User Space, Driver, and Firmware Dependencies
- Publicly Documented Interdependencies: Publicly documented interdependencies between GPU firmware, AMD GPU drivers, and ROCm user space software are now available.
- Independent Versioning of AMD GPU Driver: The AMD GPU Driver (amdgpu) is now versioned independently of ROCm.
Detailed Component Versions
The release notes provide detailed version updates for numerous ROCm components, including libraries, primitives, tools, and compilers.
Improved Documentation
ROCm 7.0.0 features updated AI training and inference benchmarking guides, new inference tutorials, documentation for early access GPU-accelerated libraries, and improved deep learning framework support.
Detailed changelog and download
You can view the full release notes for ROCm 7.0.0 on the linked GitHub release page below:
ROCm 7.0.0 Release
ROCm 7.0.0 release notes The release notes provide a summary of notable changes since the previous ROCm release. Release highlights Operating system, hardware, and virtualization support change...
