AMD Announces RoCM 7

Posted on Edited on

Quick Report

AMD has announced ROCm 7, a significant update to their GPU computing platform with major performance improvements for AI workloads. The new release focuses on developer enablement and includes support for Radeon GPUs and Windows. Alongside ROCm 7, AMD has also launched the AMD Developer Cloud, providing developers with instant access to high-performance AMD Instinct MI300X GPUs without requiring hardware investment.

Key Highlights:

  • Performance Boost: ROCm 7 delivers over 3.5x improvement in inference performance and 3x improvement in training performance compared to ROCm 6
  • Distributed Inference: New capabilities for distributed inference through collaborations with open-source projects like SGLang, vLLM, and llm-d
  • Enterprise Solutions: Introduction of ROCm Enterprise AI, an MLOps platform designed for seamless AI operations in enterprise settings
  • Expanded Support: ROCm 7 brings support for Radeon GPUs and Windows, making it available on laptops and workstations (expected in 2H 2025)
  • Developer Cloud: Launch of AMD Developer Cloud providing instant access to AMD Instinct MI300X GPUs with pre-configured environments
  • Complimentary Credits: Developers can apply for 25 free hours of compute time, with up to 50 additional hours available through programs like the ROCm Star Developer Certificate
  • Day-0 Support: Seeding MI350 systems at launch with ecosystem support including vLLM, SGLang, and pre-installed containers for popular AI frameworks

Ecosystem Growth:

AMD's AI ecosystem continues to expand with notable partnerships including:

  • Meta: Running ranking, recommendation, and content generation workloads on AMD Instinct GPUs, including Llama models
  • Microsoft: Deploying both proprietary and open-source models in production on Azure using Instinct MI300X
  • Red Hat OpenShift AI: Enabling scalable LLM inference with ROCm on Red Hat OpenShift AI for hybrid cloud environments
  • Cohere: Deploying the 104B-parameter Command R+ model on AMD Instinct GPUs using vLLM and ROCm
  • OpenAI, Midjourney, and Hugging Face: Working with AMD to accelerate inference and training workflows

Availability:

ROCm 7 will be generally available in Q3 2025, while the AMD Developer Cloud is available now to the global developer and open-source communities.

Source(s)

  • AMD Blog
  • TPU