AMD launched Instinct MI400 series GPUs and Helios rack-scale AI platform at Advancing AI 2026, featuring MI455X with HBM4 memory and claiming 15% higher compute density vs. Nvidia Vera Rubin NVL72.
AMD has unveiled a broad set of hardware and software products targeting AI infrastructure at its Advancing AI 2026 event, spanning new data centre GPUs, server processors, a rack-scale compute platform, developer tools and embedded processors for robotics. Together, these launches represent AMD's attempt to present a full-stack alternative to Nvidia across AI training, inference, networking and edge robotics.
The Instinct MI400 Series comprises two chips: the MI455X, aimed at large-scale AI inference, frontier model training and fine-tuning in AI factory environments, and the MI430X, targeting sovereign AI deployments and scientific computing workloads. Both use HBM4 memory and include security features such as secure boot, encrypted GPU-to-GPU links and hardware-based protections. The MI430X delivers up to 288 TFLOPS of hardware-based FP64 performance for scientific workloads.
"The AMD Instinct MI400 Series extends our AI portfolio with purpose-built solutions optimised for the full spectrum of AI and HPC deployments, built on an open software foundation that gives customers the flexibility to innovate at every scale," said Vamsi Boppana, Senior Vice President of AI at AMD. The GPUs run on AMD's ROCm open software stack, which provides programming models, compilers, libraries, runtimes and deployment tools designed to support interoperability and portability across different hardware and software configurations.
Helios, AMD's new rack-scale AI infrastructure platform, combines Instinct MI455X GPUs, 6th Gen EPYC server CPUs, ROCm software and AMD Pensando networking in a unified architecture designed to scale from a single rack to gigawatt-scale AI clusters. Each rack integrates 18 open rack wide-aligned compute trays containing four GPUs each, for a total of 72 GPUs per rack. A single Helios rack provides up to 2.9 exaflops of peak FP4 performance, 1.4 exaflops of peak FP8, 31 terabytes of HBM4 memory and 1.7 petabytes per second of memory bandwidth.
AMD claims the platform delivers 15 per cent more peak FP4 performance, 50 per cent more high-bandwidth memory capacity, 6 per cent more HBM bandwidth and 50 per cent more scale-out bandwidth compared to the Nvidia Vera Rubin NVL72 rack. The company also claims up to 30 per cent more tokens per dollar based on modelling using the Kimi K2 Thinking workload. "AMD Helios brings together compute, high-performance networking and open software in a unified rack-scale platform," Boppana remarked. "Together, these technologies give customers the flexibility to accelerate large-scale inference, frontier-model training and the next generation of AI infrastructure."
On the networking side, Helios uses UALink over Ethernet for scale-up connectivity and standards-based Ethernet aligned with the Ultra Ethernet Consortium for scale-out. AMD Pensando Vulcano 800 AI NICs provide high-bandwidth, low-latency connectivity for distributed inference and training workloads.
The EPYC 9006 Series represents AMD's 6th Gen server CPUs, purpose-built for agentic AI, cloud, enterprise and HPC workloads. AMD frames the modern data centre as shifting towards a three-tier architecture driven by agentic AI, with each tier creating distinct CPU requirements: agent sandbox execution, AI host nodes and general-purpose servers.
"As agentic AI reshapes how enterprise infrastructure is designed and deployed, EPYC is evolving with our customers once again," said Dan McNamara, Senior Vice President and General Manager of Compute and Enterprise AI at AMD. "6th Gen AMD EPYC delivers workload-optimised compute for the full spectrum of modern data centre needs, from running more agents and feeding GPUs across rack-scale AI infrastructure to powering the business-critical applications and services organisations rely on every day."
The EPYC 9006 lineup includes four product categories. The SP7 targets high-density agent sandbox execution with up to 256 cores and 512 threads, and in high-frequency configurations with up to 5GHz boost clocks and PCIe Gen 6 connectivity, can also serve as an AI host node. The SP8 offers 8- to 128-core options for enterprise workloads and agent sandbox execution where power efficiency is the priority, covering edge deployments, power-constrained racks and general-purpose servers. The 9006X SP7 targets HPC, technical computing and data-intensive workloads, with variants reaching up to 5.15GHz and three times the L3 cache per core of the standard SP7, aimed at simulation, modelling, large-scale analytics and in-memory workloads. The 9006 LP—formerly codenamed Verano—is built specifically as an AI host node CPU for rack-scale AI systems such as Helios, using LPDDR memory and up to 5GHz frequency to keep accelerators fed across dense, GPU-rich racks. All four share a common architecture, software ecosystem and security foundation, allowing customers to mix and match processors across different data centre roles.
AMD has announced ROCm.ai, a new AI-native software experience for developers working across AMD platforms. It combines AI-assisted development, intelligent deployment and AI-powered optimisation into a unified toolset with three core components. ROCm CLI is a new command-line interface for installing, validating, serving and managing AI workloads on AMD hardware, including hardware-aware workflows, automated environment setup and support for air-gapped deployments. AMD Skills integrates AMD-authored expertise into AI coding assistants including Claude, Cursor and Codex, providing AMD-specific recommendations for installing, migrating, debugging and optimising AI workloads. Hyperloom is a new open-source, agentic system that automates the optimisation of end-to-end inference workloads, handling profiling, analysis, kernel optimisation and validation.
"ROCm.ai brings agentic AI development to AMD platforms—agents that don't just answer questions, but profile, debug and drive workloads toward peak performance on AMD hardware," said Anush Elangovan, Corporate Vice President of AI Software at AMD. "Whether standing up an environment, serving a model or closing the gap to peak performance, ROCm.ai does the heavy lifting—right inside the coding assistants developers already use."
AMD claims that AI-powered optimisations delivered through ROCm.ai produce an average 3.3x inference improvement and 2.4x training improvement for the latest version of ROCm over ROCm 7 on the same hardware. These figures were measured on Instinct MI355X GPU systems running GLM-5, Kimi-K2.5 and DeepSeek-R1-0528 models for inference, and DeepSeek-V2-Lite, DeepSeek-V3-16B and Qwen3-30B-A3B models for training. ROCm.ai availability is expected to begin in August 2026.
AMD, AT&T and Microsoft have announced OTel 2.0, an open-source AI model trained specifically for the telecommunications industry. AT&T has processed more than 1 trillion tokens through managed compute in Microsoft Foundry, using AMD Instinct GPUs to train the OTel 2.0 models. The telco giant processes an average of 45 billion tokens daily and uses an intelligent gateway to match each task with an appropriate model. AT&T reports that its cache-aware model-routing approach has cut AI costs by as much as 80 per cent and saved millions of dollars. The work builds on the Open Telco AI initiative launched at MWC 2026 by AMD, AT&T, GSMA and other industry participants, which aims to provide the telecommunications industry with AI models trained on domain-specific data.
"By combining AMD Instinct GPUs, the open AMD ROCm software platform and flexible compute from core to edge, we can help operators build and deploy telco-specific AI at scale while balancing accuracy, performance and cost," McNamara added.