AMD has launched Helios, a fully integrated rackscale AI infrastructure solution designed to compete directly with proprietary AI systems by combining its latest Instinct GPUs, EPYC server processors, and Pensando networking components using open industry standards. The system integrates 72 AMD Instinct MI455X GPUs with AMD EPYC "Venice" CPUs and AMD Pensando "Vulcano" networking, delivering 2.9 exaFLOPS of FP4 compute performance and 1.4 exaFLOPS of FP8 performance for both frontier model training and large-scale inference workloads.

What Makes AMD's Helios Different From Competitors?

The Helios design prioritizes openness and modularity, built on industry standards including OCP Open Rack Wide (ORW), Ultra Accelerator Link (UALink), and Ultra Ethernet Consortium (UEC) specifications. This approach contrasts with proprietary rack designs that lock customers into single-vendor ecosystems. The system delivers 31 terabytes of HBM4 memory capacity with 19.6 terabytes per second of memory bandwidth per GPU, enabling larger language models, extended context windows, and multi-agent workflows that require rapid data access.

The rack's architecture emphasizes both scale-up and scale-out connectivity. Four scale-up cartridges use UALink over Ethernet to connect up to 72 GPUs with 260 terabytes per second of aggregate bandwidth, while the scale-out design delivers 43 terabytes per second using standard Ethernet-based Pensando networking. AMD claims this represents over 50 percent more scale-out bandwidth than competing solutions.

How Does Helios Support Enterprise AI Deployment?

• Security Architecture: The system includes hardware-enforced isolation, device identity verification, encrypted memory, and encrypted interconnects, enabling secure multi-tenant AI workloads at hyperscale with continuous attestation at every layer.
• Serviceability Design: Modular construction with integrated power, cooling, and networking eliminates recabling during maintenance, enabling rapid sled replacement and maximizing uptime for mission-critical AI operations.
• Software Ecosystem: AMD ROCm software translates the rackscale architecture into production-ready performance, natively supporting PyTorch, TensorFlow, and JAX frameworks while enabling fleet-scale deployment and lifecycle management.

Who Is Adopting AMD's New Infrastructure?

AMD has announced multiple strategic partnerships to deploy Helios at scale. The company revealed collaborations with OpenAI and Meta, each committing to deploy 6 gigawatts of AMD GPU capacity. Additional partnerships include expansions with Oracle to help customers achieve next-generation AI scale, collaboration with Tata Consultancy Services (TCS) to bring Helios architecture to India, and expanded work with Hewlett Packard Enterprise (HPE) to advance open RackScale AI infrastructure.

The partnerships extend beyond hyperscalers. AMD announced collaborations with Celestica to advance the next era of AI infrastructure, and with Nutanix to develop an open, scalable platform for enterprise AI. These partnerships signal AMD's strategy to position Helios as the infrastructure backbone for organizations seeking sovereign AI capabilities and vendor independence.

What Technical Advantages Does Helios Offer for AI Workloads?

The MI455X GPUs at the heart of Helios use AMD's next-generation CDNA 5 architecture, delivering massive memory capacity and extreme bandwidth optimized for both training and inference. The system's 31 terabytes of total HBM4 memory enables researchers and enterprises to run larger models without splitting them across multiple racks, reducing complexity and latency in distributed systems.

The EPYC "Venice" CPUs provide up to 256 high-performance cores with 1.6 terabytes per second of memory bandwidth, serving as host processors that sustain rackscale performance. AMD Pensando AI NICs enable large-scale, scale-out AI networking with fully programmable hardware, UEC-ready remote direct memory access (RDMA), and high-performance Ethernet for intelligent traffic management and advanced congestion control across racks.

The design also incorporates AMD Pensando Salina data processing units (DPUs) with P4-programmable, line-rate networking and 16 Arm N1 cores. These DPUs offload networking, storage, and security services from the main compute path, freeing GPU resources for actual AI workloads rather than infrastructure overhead.

Why Does Open Standards Matter for AI Infrastructure?

By designing Helios on open standards across compute, networking, rack design, and software, AMD aims to give customers choice, flexibility, and long-term control over their AI infrastructure investments. This approach contrasts with closed ecosystems where customers become dependent on a single vendor for hardware, software, and support. The use of OCP standards, UALink, and UEC specifications means organizations can mix and match components from multiple vendors while maintaining compatibility.

The ROCm software stack further reinforces this openness, providing transparent, community-driven AI development with native support for leading frameworks. This enables developers to preserve familiar workflows while maximizing performance for frontier-scale workloads, reducing vendor lock-in and accelerating time-to-production for AI applications.