At the Advancing AI 2025 event in San Jose, AMD pulled the curtain back on its most ambitious AI strategy yet—delivering a sweeping portfolio of hardware, software, and full-stack infrastructure to power an open, integrated, and scalable AI future.
Key Highlights:
-
Instinct MI350 Series GPUs launched with 4x AI performance gen-on-gen, delivering a massive leap in compute and efficiency.
-
Helios, AMD’s next-gen rack-scale AI platform, was unveiled, targeting 10x performance for inference on future MI400 GPUs and Zen 6 CPUs.
-
AMD reinforced its commitment to open AI, introducing ROCm 7.0 and a global Developer Cloud Access Program.
-
Industry giants—including Meta, Microsoft, OpenAI, Oracle, xAI, Cohere, HUMAIN, and others—showcased real-world deployments and future collaborations with AMD AI.
AMD Chair and CEO Dr. Lisa Su, joined on stage by partners like Meta, Microsoft, Oracle, and OpenAI, declared:
“We are entering the next phase of AI, driven by open standards, shared innovation, and AMD’s expanding leadership across a broad ecosystem.”
New AI Muscle: Instinct MI350 Series
The new Instinct MI350X and MI355X GPUs bring:
-
4x AI compute uplift over the previous generation¹
-
35x increase in inferencing performance²
-
Up to 40% more tokens-per-dollar³ vs. competitors
-
Deployed already in Oracle Cloud Infrastructure (OCI) and targeting wider rollout in 2H 2025
ROCm 7.0: AI Software, Reinvented
AMD’s open-source stack now supports:
-
Broader framework compatibility
-
Expanded hardware support
-
Enhanced tools, APIs, and libraries for developers
Additionally, the AMD Developer Cloud is now widely available, purpose-built for high-performance AI development with ROCm 7 pre-installed.
Rack-Scale Innovation: “Helios” and Beyond
AMD previewed its Helios rack platform, coming in 2026:
-
Built on Instinct MI400 GPUs, Zen 6 “Venice” EPYC CPUs, and Pensando “Vulcano” NICs
-
Designed to deliver 10x inference performance on complex Mixture-of-Experts models⁴
And that’s just the beginning—AMD set a new 2030 sustainability goal:
➡️ 20x increase in rack-scale energy efficiency⁵, slashing AI training rack needs from 275 to just one⁶.
Ecosystem Power Moves
-
Meta uses MI300X for Llama 3 & 4 inference, with plans for MI400.
-
OpenAI runs GPT models on MI300X in Azure and is working closely on MI400.
-
Oracle to deploy 131,072 MI355X GPUs in new zettascale clusters.
-
Microsoft runs both proprietary and open models using MI300X on Azure.
-
Cohere, HUMAIN, Red Hat, Astera Labs, and Marvell showcased AMD-powered AI breakthroughs.
From next-gen chips to open software and collaborative AI infrastructure, AMD is going all-in to rewrite the rules of the AI race—with openness, scale, and performance at the core.
