AMD Unveils Helios: A Rack-Scale AI Platform with 50% More Memory Than NVIDIA's Vera Rubin

At the OCP Global Summit 2025, AMD took the stage to introduce its latest rack-scale AI hardware platform, Helios. This new offering promises significant improvements in serviceability and memory capacity, positioning itself as a strong competitor to NVIDIA’s Vera Rubin platform. For data center operators and AI practitioners, these advancements could translate into more efficient and scalable deployments.

What Changed Technically?

Helios is designed with several key technical innovations that address common pain points in rack-scale AI infrastructure:

• 50% More Memory: Helios boasts 128 GB of HBM3 memory per GPU, compared to the 84 GB found in NVIDIA’s Vera Rubin. This increase in memory capacity allows for handling larger models and datasets without the need for frequent data swapping or distributed training setups.

• Easier Serviceability: AMD has introduced a modular design that simplifies hardware maintenance. Key components like GPUs, network interfaces, and storage can be swapped out with minimal downtime, reducing the total cost of ownership (TCO) over the platform’s lifecycle.

• Scalable Architecture: Helios supports up to 16 GPUs in a single rack unit (2U), enabling high-density configurations that maximize compute power per rack space. This is crucial for data centers looking to optimize their footprint and energy efficiency.

Why It Matters

For AI practitioners, the increased memory capacity means more flexibility in model training and inference:

• Larger Models: With 128 GB of HBM3 memory, Helios can support larger models that require more parameters. This is particularly beneficial for cutting-edge applications like natural language processing (NLP) and computer vision.

• Efficient Data Handling: The additional memory reduces the need for data offloading to external storage or distributed training setups, leading to faster training times and lower latency in inference.

Implementation Details

Let’s dive into some of the technical specifics:

• Memory Configuration:

  • Each GPU is equipped with 128 GB of HBM3 memory.
  • The platform supports up to 16 GPUs per 2U rack unit, providing a total of 2 TB of memory in a single rack.

• Interconnects:

  • Helios uses high-speed interconnects (like PCIe 5.0 and CXL 2.0) to ensure low-latency communication between GPUs and other system components.
  • This is essential for maintaining performance in multi-GPU setups and distributed training environments.

• Cooling and Power Efficiency:

  • AMD has integrated advanced liquid cooling solutions to manage the thermal load of high-performance GPUs.
  • The platform is designed to be energy-efficient, with a focus on reducing power consumption without compromising performance.

Benchmarks

While specific benchmark results were not provided at the OCP Global Summit, early tests suggest that Helios can handle complex AI workloads more efficiently than its predecessors. The combination of increased memory, improved serviceability, and scalable architecture makes it a compelling choice for data centers looking to future-proof their infrastructure.

Conclusion

AMD’s Helios platform represents a significant step forward in rack-scale AI hardware. With 50% more memory than NVIDIA’s Vera Rubin and a modular design that simplifies maintenance, it offers data center operators and AI practitioners the tools they need to tackle larger and more complex models with greater efficiency.