AMD releases ROCm 7 with performance improvements and support for AI models

AMD announced at the Advancing AI 2025 event the launch of ROCm 7, the latest version of its software platform for developing and running artificial intelligence models.

This release incorporates performance improvements, improved compatibility with existing models, and enterprise capabilities for large-scale deployment.

ROCm 7 includes optimized support for models such as Llama 3, DeepSeek R1, and Qwen2-72B. According to internal testing, this new version achieves up to 3.8 times better performance on inference tasks compared to the previous version, ROCm 6.

The platform also supports the latest open-source frameworks such as PyTorch, JAX, vLLM, and MaxText, and supports advanced data types such as FP4, FP6, FP8, and BF16 to improve training and inference efficiency.

Enterprise capabilities and data center deployment

ROCm 7 incorporates tools for workload management, integration with Kubernetes, and support for Linux and Windows operating systems (via WSL).

These features allow the platform to be adapted to data center production environments, facilitating the adoption of AI models at scale.

AMD continues its focus on open source software, offering frequent updates, developer documentation, and support for distributed models.