Leadership Rackscale Performance for Frontier AI
The AMD Helios Rackscale solution design is fully integrated AI infrastructure, combining the latest AMD Instinct™ GPUs, AMD EPYC™ Server CPUs, AMD Pensando™ networking, designed using open industry standards enabling large-scale inference, frontier model training and fine-tuning.
News on Customer and Partner Adoption
AMD Helios Rackscale Highlights
The AMD Helios rackscale design integrates 72 AMD Instinct™ MI455X GPUs with AMD EPYC™ “Venice” CPUs and AMD Pensando™ “Vulcano” networking using UALink, forming a tightly co-designed architecture optimized for compute, data movement, and system efficiency
AMD Helios delivers industry-leading memory capacity and bandwidth enabling larger models, context windows and super responsive multi-agent workflows. Designed on open standards including OCP Open Rack Wide (ORW), Ultra Accelerator Link (UALink™), and Ultra Ethernet Consortium (UEC), AMD Helios scales efficiently across datacenters while optimizing power, cooling, and serviceability for modern AI infrastructure.
AMD Helios Rackscale Solution Specs
Key Rack Features
Supporting both OCP and MX datatypes delivering 2.9EF3 of FP4 and 1.4 EF3 of FP8 compute for AI inference and training needs
Industry leading 31TB of latest generation HBM4 offering 19.6 TB/s of memory bandwidths
Four scale-up cartridges leverage UALink™ over Ethernet (UALoE) to seamlessly connect up to 72 GPUs per AMD Helios Rackscale Solution with up to an industry leading 260 TB/s aggregate scale-up bandwidth.
Delivering over 50% more than competition with 43 TB/s of scale out bandwidth using industry standard ethernet based Pensando networking3
Acts as a centralized power hub to manage and distribute power across the rack via a vertical busbar
Distributes liquid coolant to the compute and switch trays via quick-disconnect connections, enabling efficient heat dissipation
AMD Helios secures AI at scale—protecting data, models, and workloads with hardware root of trust and continuous attestation at every layer. AMD Helios supports hardware-enforced isolation, device identity, encrypted memory and interconnects. The result is a secure-by-design platform design that enables trusted, multi-tenant AI at hyperscale.
The AMD Helios rack design is purpose-built for rack-scale serviceability, combining a modular design with infrastructure optimized for fast, low-disruption maintenance. Integrated power, cooling, and networking connections eliminate recabling, enabling rapid, seamless sled replacement. The result is resilient, always-on infrastructure that maximizes uptime at hyperscale.
AMD Helios is designed on open standards across compute, networking, rack design and software—enabling choice, flexibility, and long-term control.
AMD Helios Rackscale Solutions Compute Tray
-
AMD EPYC™ CPUAMD EPYC™ "Venice" CPU powered by groundbreaking "Zen 6" architecture, with up to 256 high performance cores, and 1.6TB/s memory bandwidth -
AMD Instinct™ GPUAMD Instinct™ MI455X GPUs x4 using next generation AMD CDNA™ 5 architecture with up to 432GB HBM4 memory and 19.6 TB/s memory bandwidth per GPU -
AMD Pensando™ AI NICAMD Pensando™ "Vulcano" AI NICs with 800 Gbps for AI scale out bandwidth, PCIe® Gen 6 and OCP form factors, and UALink for direct connectivity to CPUs and GPU -
AMD Infinity Fabric™ TechnologyHigh-performance communication fabric between CPU and GPU -
AMD Pensando™ DPUAMD Pensando™ Salina DPU with P4‑programmable, line‑rate networking and 16 Arm® N1 cores, offloading networking, storage, and security services for front‑end server‑to‑client connectivity.
AMD Instinct™ MI455X GPUs
Built on the next-generation AMD CDNA™ 5 architecture, delivering massive HBM4 memory capacity, extreme bandwidth, and leading AI compute performance —enabling frontier-class training, distributed inference, and efficient scaling across the AMD Helios Rackscale design.
6th Gen AMD EPYC™ CPUs
Serves as the host processors for AMD Helios, delivering high core counts, exceptional memory bandwidth, and enterprise-grade reliability to sustain rackscale performance.
AMD Pensando™ AI NICs
AMD Pensando™ AI NICs enable large-scale, scale-out AI networking with fully programmable hardware and software, UEC-ready RDMA, and high-performance Ethernet—delivering intelligent traffic management, advanced congestion control, and high-throughput inter-rack connectivity at rack scale.
Open, Scalable AI Software for AMD Helios Infrastructure
AMD ROCm™ software powers the AMD Helios platform design by translating its rackscale architecture into production-ready AI performance. Built for transparent, community-driven AI development, ROCm natively supports leading frameworks including PyTorch, TensorFlow, and JAX, enabling high-throughput inference and efficient distributed training while preserving familiar developer workflows.
It supports fleet-scale deployment, observability, and lifecycle control for AI environments. With optimized libraries, open standards, and broad framework compatibility, ROCm helps ensure portability while maximizing performance for frontier-scale workloads.
Frequently Asked Questions
The AMD Helios Rackscale solution is the first rackscale AI reference design from AMD, built fully on Meta’s new Open Rack Wide (ORW) standard submitted to the Open Compute Project (OCP). It combines 72 AMD Instinct MI455X GPUs, AMD EPYC™ CPUs, and AMD Pensando™ DPUs along with other leading components in an open, double-wide ORW rack designed for today’s AI-scale datacenters.
At OCP Global Summit, Meta introduced a new Open Rack for AI featuring the ORW form factor. AMD is aligning with this open design through the AMD Helios Rackscale solution, showing the first full rack built on ORW standards. It demonstrates the AMD commitment to openness — from silicon to system to rack to large-scale clusters.
Earlier OCP racks were single-wide, focused on general-purpose servers. ORW introduces a double-wide rack optimized for high-density AI systems, providing higher power, liquid cooling, and wider trays to accommodate large accelerator nodes. The AMD Helios Rackscale solution leverages ORW to integrate GPUs, CPUs, DPUs, and other components into a unified, serviceable, easily deployable AI system.
A full AMD Helios Rackscale solution delivers up to 1.4 exaFLOPS of FP8 compute, 2.9 exaFLOPS of FP4 compute and 31 TB of HBM4 memory, enabling trillion-parameter AI training and large-scale AI inference.
Each AMD Instinct MI455X GPU uses next-gen the AMD CDNA™ architecture, with 432 GB HBM4 of memory and up to 19.6 TB/s bandwidth — ideal for Frontier-class AI models inference and training. Each GPU delivers up to 40 PFLOPS FP4 and 2.9 EF FP4 per AMD Helios Rackscale solution.
The AMD Helios Rackscale solution is the first double-wide ORW rack solution built for real-world operations, combining modular trays and easy serviceability with sustained performance in power-constrained datacenters. It reduces fragmentation by replacing proprietary, one-off designs with open standards — including OCP, UALink, and UEC — enabling efficient, high-performance AI clusters at scale.
It’s a reference design, not a product for sale. The AMD Helios Rackscale solution is a blueprint for OEM and ODM partners to build their own branded systems based on open ORW standards
The platform is powered by AMD ROCm™, an open-source AI and HPC software stack with Day-0 support for PyTorch, TensorFlow, JAX, ONNX Runtime, vLLM, and Triton. It scales from a single GPU to multi-rack superclusters and avoids vendor lock-in.
The AMD Helios Rackscale solution reference design is being shared with partners now, with volume deployments expected in 2H 2026.
Stay Informed
Sign up to receive the latest data center news and server content.
Footnotes
- Based on AMD internal analysis, actual results subject to change. MI450-001
- Based on AMD internal analysis, actual results subject to change. MI450-001