Semiconductor company AMD is expanding beyond chips and putting software at the center of its artificial intelligence (AI) strategy, to make it easier for developers and enterprises to build and scale AI models.
“AI performance and adoption are no longer determined by silicon alone,” said Anush Elangovan, VP of AI Software at AMD, told Back End News in an email interview. “Hardware leadership remains our foundation, but software defines how quickly that performance translates into value.”
At the center of this shift is AMD’s ROCm open software platform, which integrates CPUs, GPUs, NPUs, and adaptive computing into a single environment. The company said ROCm now supports workloads ranging from small model tuning to large-scale AI training, including models with up to 520 billion parameters on a single GPU, without locking users into proprietary systems.
The latest update, ROCm 7, adds features designed to address one of the biggest challenges in AI today: scaling models efficiently. The platform introduces distributed inference, allowing AI models to run across multiple GPUs or clusters instead of a single device. This enables organizations to deploy large models, including reasoning and agent-based systems, more efficiently.
ROCm 7 also supports lower-precision data formats such as FP4 and FP6, reducing memory use and power consumption while maintaining accuracy. AMD said these improvements deliver up to 4.8x faster inference and three times faster training compared with the previous version.
For enterprises, the impact is direct: faster model training, lower infrastructure costs, and better energy efficiency. Because ROCm is open-source, developers can modify and integrate it into their own systems across cloud, on-premises, or edge environments.
“ROCm software is fully open-source, allowing users to inspect, customize, and optimize the stack to fit their specific workloads,” Elangovan said.
The platform supports widely used AI frameworks such as PyTorch, TensorFlow, and JAX, and integrates with tools like vLLM. AMD also said more than 2.1 million models from Hugging Face can run on ROCm with minimal changes. These partnerships aim to reduce friction for developers.
“Innovation happens faster when developers are empowered and the community builds together,” Elangovan said.
To further simplify access to AI infrastructure, AMD introduced the AMD Developer Cloud, which provides on-demand access to Instinct MI300X GPUs in pre-configured environments. Developers can log in, launch a workspace, and start training models within minutes without setting up hardware or software. The company said the service includes free GPU hours for eligible users, lowering the barrier to entry for startups and individual developers.
As AI workloads grow more complex, AMD identified three main challenges: scalability, efficiency, and integration. ROCm 7 addresses these through features like distributed inference and Mixture-of-Experts scheduling, which improve how workloads are managed across multiple GPUs.
AMD also highlighted its hardware portfolio, including Instinct GPUs and EPYC CPUs, as part of a unified AI stack designed to improve performance per watt and reduce total cost of ownership.
“Software turns our hardware leadership into strategic differentiation,” Elangovan said.
AMD plans to accelerate ROCm updates, expand ecosystem partnerships, and build energy-efficient AI infrastructure, including rack-scale systems. The company is targeting long-term gains such as a 20x improvement in energy efficiency by 2030, as it competes to provide end-to-end AI solutions for enterprises and developers.
You must be logged in to post a comment.