AI infrastructure is being redefined at the scale of the entire rack. In this blog, you will explore how the AMD CDNA™ 5 architecture, the AMD Instinct™ MI455X GPU, and the AMD Helios rackscale solution come together as a single open platform for large-scale AI training and inference. You will see how AMD balances compute, memory, and networking across 72 GPUs, how open standards keep the platform flexible, and how the AMD ROCm™ software stack ties it all into a complete, deployable system, from silicon to rack. By the end, you will understand what changes when AI infrastructure is designed as a unified rack rather than a collection of accelerators, and where AMD is taking it next.
AI Infrastructure Must Evolve Beyond the GPU# #
Artificial intelligence has become one of the largest drivers of computing innovation in history. Foundation models are rapidly growing, context windows continue to expand, and AI factories are driving the next AI infrastructure shift.
As models have grown, simply building faster GPUs is no longer enough. Performance now depends on the entire system: compute, memory, networking, software, power, cooling, and serviceability all working together as a unified platform.
AMD addresses this need with an integrated AI solution that combines the AMD CDNA™ 5 architecture, AMD Instinct™ MI455X GPUs, AMD EPYC™ “Venice” processors, AMD Pensando™ networking, ROCm™ software, and the AMD Helios rackscale platform. Together they deliver an open, standards-based platform designed for rack-scale AI infrastructure.
Reimagining AI Compute# #
At the heart of Helios is the AMD Instinct™ MI455X GPU, designed specifically for the next generation of AI training and inference.
Built on the AMD CDNA™ 5 architecture (shown in Figure 1, below), MI455X introduces significant advances across every aspect of the GPU.
Up to
4x greater AI compute throughput for key low-precision formatsMI400-006432 GB HBM4 memory2.91x memory bandwidth MI400-008Larger caches and local memories
Improved utilization through architectural enhancements
Rather than focusing on a single specification, CDNA™ 5 improves the balance between compute, memory, and communication — the three resources that increasingly determine AI performance.
Memory Designed for Large Models# #
Modern AI workloads increasingly rely on high-bandwidth memory and massively parallel compute, in addition to traditional CPU-based compute, to deliver new capabilities and greater model efficiency.
CDNA™ 5 expands this horizon with a substantially redesigned memory hierarchy.
HBM4 increases both capacity and bandwidth to support the largest models and context windows, while serving more concurrent inference sessions.
Inside the GPU, larger L2 caches, configurable local memory, Tensor Data Movers, and expanded register files optimize data movement and keep more information closer to the compute it needs.
The result is an accelerator capable of supporting the most powerful AI engines without becoming memory-bound.
From Eight GPUs to an Entire Rack# #
Perhaps the largest architectural shift is how AMD approaches scaling.
Previous GPU generations typically operated in groups of eight accelerators.
Built on Open Compute Project (OCP) Open Rack Wide (ORW) standards, Helios transforms 72 GPUs into a unified AI system while simplifying deployment, servicing, and future infrastructure evolution, as shown in Figure 2, below.
Built on Open Networking# #
As AI factories continue to grow, scale-out networking becomes just as important as GPU performance.
AMD Helios embraces an open networking strategy based on industry-standard Ethernet, delivering up to 43 TB/s scale-out bandwidth.
Support for AMD Pensando™ AI-NICs and open Ethernet networking enables customers to scale efficiently from a single rack to large AI clusters. The platform leverages Ultra Accelerator Link™ over Ethernet (UALoE) for high-bandwidth rack-scale communication, delivering 260 TB/s of scale-up bandwidth. It also uses Ultra Ethernet Consortium (UEC) technologies for scale-out networking and an open Ethernet switching ecosystem based on the Open ESUN framework.
Rather than requiring proprietary fabrics, AMD enables customers to leverage an open ecosystem of switches, NICs, and management software.
AI Infrastructure Beyond Silicon# #
AMD Helios is more than a collection of GPUs.
Each rack integrates:
72 AMD Instinct™ MI455X GPUs
31 TB of HBM4 GPU memory18 AMD EPYC™ “Venice” processors, each with up to 256 coresUp to 36 TB of DDR5 memory****260 TB/s high-bandwidth scale-up switchingOpen Ethernet scale-out networking
Rack-level management
Direct liquid cooling
Tray serviceability, observability, and telemetry
Wave32 compute execution
The platform is designed not only for performance, but also for deployment, operation, and maintenance at hyperscale.
Integrated telemetry, health monitoring, power and thermal management, and modular serviceability help operators maintain availability while reducing operational complexity.
Software Completes the Solution# #
Hardware alone is not enough. The AMD ROCm™ software stack enables developers to take advantage of the CDNA™ 5 architecture with minimal effort, while providing an open environment for AI frameworks, communication libraries, runtime software, and developer tools.
Together, ROCm™, AMD Instinct™ MI455X, and AMD Helios deliver a complete AI platform, from silicon to rack.
Looking Ahead# #
As AI infrastructure evolves from individual accelerators into complete rack-scale systems, architectural innovation must extend beyond the GPU itself. The AMD CDNA™ 5 architecture, AMD Instinct™ MI455X GPU, and AMD Helios rackscale solution represent the next step by AMD toward building open, scalable AI infrastructure capable of supporting frontier-model training, large-scale inference, and the next generation of AI factories.
Expect this platform to keep expanding as the open ecosystem around it grows, with more open networking options, deeper ROCm™ integration, and additional rack-scale configurations. Follow the ROCm blog for architectural deep dives and hands-on guides as Helios moves from introduction to deployment.
Summary# #
In this blog, you explored how AMD unifies the AMD CDNA™ 5 architecture, the AMD Instinct™ MI455X GPU, AMD EPYC™ “Venice” processors, AMD Pensando™ networking, and the AMD ROCm™ software stack into the AMD Helios rackscale solution, a single open platform that scales from eight accelerators to 72 GPUs in one rack. You saw how CDNA™ 5 rebalances compute, memory, and communication; how HBM4 and a redesigned memory hierarchy keep the largest models from becoming memory-bound; and how open, Ethernet-based scale-up and scale-out networking let customers grow without proprietary fabrics. Together, these technologies deliver a complete AI platform, from silicon to rack.
To go deeper into the compute, memory, networking, software, and rack design behind Helios, use the resources below, and check back on the ROCm blog for follow-on posts as the platform rolls out.
## Additional Resources[#](#additional-resources)
## Endnotes[#](#endnotes)
The performance claims in this blog reference the AMD endnotes below.
MI400-006— Based on AMD Performance Labs calculations (June 2026) using an AMD Instinct MI455X GPU, peak theoretical precision performance (FP32, FP16, BF16, MXFP6, MXFP8, FP8, MXFP4 Matrix/Vector), compared to published specifications for AMD Instinct MI355X, MI350X, MI325X, MI300X, MI250X, and MI100 GPUs. Results may vary by system configuration and datatype.MI400-008— Calculations by AMD Performance Labs in June 2026, based on the published memory capacity and memory bandwidth specifications of an AMD Instinct MI455X GPU vs the published memory capacity and memory bandwidth specifications of AMD Instinct MI355X, MI350X, MI325X, MI300X, MI250X and MI100 GPUs, respectively. System manufacturers may vary configurations, yielding different results.
Disclaimers# #
The information presented in this document is for informational purposes only and may contain technical inaccuracies, omissions, and typographical errors. The information contained herein is subject to change and may be rendered inaccurate for many reasons, including but not limited to product and roadmap changes, component and motherboard version changes, new model and/or product releases, product differences between differing manufacturers, software changes, BIOS flashes, firmware upgrades, or the like. Any computer system has risks of security vulnerabilities that cannot be completely prevented or mitigated. AMD assumes no obligation to update or otherwise correct or revise this information. However, AMD reserves the right to revise this information and to make changes from time to time to the content hereof without obligation of AMD to notify any person of such revisions or changes.
THIS INFORMATION IS PROVIDED “AS IS.” AMD MAKES NO REPRESENTATIONS OR WARRANTIES WITH RESPECT TO THE CONTENTS HEREOF AND ASSUMES NO RESPONSIBILITY FOR ANY INACCURACIES, ERRORS, OR OMISSIONS THAT MAY APPEAR IN THIS INFORMATION. AMD SPECIFICALLY DISCLAIMS ANY IMPLIED WARRANTIES OF NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR ANY PARTICULAR PURPOSE. IN NO EVENT WILL AMD BE LIABLE TO ANY PERSON FOR ANY RELIANCE, DIRECT, INDIRECT, SPECIAL, OR OTHER CONSEQUENTIAL DAMAGES ARISING FROM THE USE OF ANY INFORMATION CONTAINED HEREIN, EVEN IF AMD IS EXPRESSLY ADVISED OF THE POSSIBILITY OF SUCH DAMAGES.
AMD, the AMD Arrow logo, and combinations thereof are trademarks of Advanced Micro Devices, Inc. Other product names used in this publication are for identification purposes only and may be trademarks of their respective companies.
© 2026 Advanced Micro Devices, Inc. All rights reserved