AMD at CES: Chipmaker unveils Helios AI rack, Ryzen AI Halo and Gene One humanoid in full-stack push from data centre to robotics

Helios rack targets next-generation AI data centres

AMD has showcased Helios, which it describes as the world’s most advanced AI rack, designed to meet the demands of next-generation data centres.

Helios uses a double-wide design based on the Open Compute Project open rack-wide standard, developed in collaboration with Meta. The rack weighs nearly 7,000 pounds and is engineered for serviceability, manufacturability and long-term reliability. AMD said each rack consumes more power than two compact cars, reflecting the scale of compute being delivered.

Compute tray combines GPUs, CPUs and networking

At the heart of Helios is the compute tray. This integrates four MI455 GPUs with the next-generation EPYC Venice CPU and Pensando networking chips. All components are liquid-cooled to maximise sustained performance and energy efficiency.

AMD introduced the MI455X GPU alongside the rack. The chip contains 320 billion transistors, around 70% more than its predecessor, and is built from 12 compute and input-output chiplets using two-nanometre and three-nanometre process technologies. Each GPU includes 432 gigabytes of high-bandwidth HBM4 memory.

EPYC Venice doubles down on AI workloads

The EPYC Venice CPU is also new. Built on a two-nanometre process, it offers up to 256 high-performance Zen 6 cores. AMD positioned Venice as a purpose-built AI CPU, with doubled memory capacity and increased GPU bandwidth compared with the previous generation.

Networking across the Helios rack is handled by Pensando Volcano and Selena chips using 800 gigabit Ethernet. This enables ultra-high bandwidth and low-latency communication, allowing tens of thousands of Helios racks to scale across a single data centre.

Each rack delivers more than 18,000 CDNA 5 GPU compute units and over 4,600 Zen 6 CPU cores. AMD said this equates to up to 2.9 exaflops of performance, supported by 31 terabytes of HBM4 memory.

Luma pushes multimodal and world simulation AI

AMD also highlighted progress from Luma, which is focused on building multimodal and general intelligence systems.

Luma’s models combine audio, video, language and images to simulate physics, causality and real-world behaviour. A demonstration showed its Ray 3 system editing and modifying both real and AI-generated footage, effectively allowing users to generate and manipulate entire worlds.

The technology is positioned as enabling hybrid human-AI production, where creators guide motion and timing while models generate content. AMD said this could allow filmmakers and designers to create cinematic environments without physical sets.

Ryzen AI 400 Series targets notebooks and creators

On the client side, AMD announced the Ryzen AI 400 Series. The chips feature up to 12 Zen 5 CPU cores, up to 16 RDNA 5 and RDNA 3.5 GPU cores and support for faster memory speeds.

The processors deliver up to 60 trillion operations per second of AI compute and are aimed at improving content creation and multitasking. AMD said notebooks based on the platform will begin launching at CES, with more than 120 ultra-thin gaming and commercial PCs expected during the year.

Ryzen AI Halo brings large models to the desktop

AMD also revealed Ryzen AI Halo, a new reference platform for local AI deployment. The compact desktop system is capable of running models with up to 200 billion parameters locally, powered by the highest-end Ryzen AI Max processor.

Ryzen AI Halo supports multiple operating systems and ships with the latest ROCm software stack and open-source developer tools. AMD said this allows developers to build, test and deploy AI applications without relying on cloud infrastructure. The platform is due to launch in the second quarter, with demos showing 3D generative models building environments from a handful of smartphone images.

Gene One humanoid robot moves towards production

Rounding out the announcements was Gene One, a humanoid robot designed for industrial and healthcare settings. Italian in design, Gene One features a distributed tactile skin across its body, allowing it to sense pressure, contact and intention.

AMD said touch is treated as a primary source of intelligence, enabling safer and more natural interaction between humans and robots. Powered by AMD compute platforms, Gene One is expected to enter commercial manufacturing in the second half of 2026. The company is already working with industrial partners, including a major steel producer, to deploy the robots in safety-critical environments.

Together, the announcements underline AMD’s strategy of extending AI capability from hyperscale data centres to personal devices and physical robots, positioning the company as a full-spectrum AI computing provider.