June 13, 2012 - In the early hours of this morning's AMD Advancing AI 2025 event, AMD officially launched itsthe next generationopen source software stack technology ROCm 7, further accelerating AI and developer productivity.
With the release of ROCm 7, AMD is finally moving forward from its ROCm 6 software stack, which has undergone several updates over the past few years -- especially since the advent of AI computing. Here are some of the features AMD is focusing on in ROCm 7:
- Latest Algorithms and Models
- Advanced Features for Extended AI
- MI350 series support
- Cluster Management
- Enterprise Features
With ROCm, AMD says it is focusing more on the growing inference capabilities in its software stack.The ROCm 7 stack will include enhanced frameworks, such as vLLM v1, llm-d, SGLang, and is focused on providing a wide range of optimizations. Upcoming new kernels and algorithms for ROCm 7 include GEMM auto-tuning, MoE, Attention, and Python-based kernel writing.
AMD has announced FP6 and FP4 support for its MI350 series, and ROCm 7 also includes full support for these advanced data types such as FP8, FP6, FP4, and mixed precision.
1AI has learned from the launch that in terms of performance, theAMD Says ROCm 7 Brings Up to 3.5x Performance Boost for AI Workloads with a Focus on Inference.
Specifically, compared to ROCm 6, ROCm 7's Llama 3.1 70B is upgraded by a factor of 3.2, and Qwen2-72B is upgraded by a factor of 3.4.DeepSeek R1 3.8x improvement.