Warning Shots Fired as AMD Announces New Data Center GPUs

2026-07-27 · Source: Big Data & AI News - EE Times · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Cloud Computing & IT Infrastructure, Emerging Technologies & Innovation · Depth: Advanced, long

Summary

AMD's Advancing AI developer conference on July 27, 2026, showcased competitive new CPUs, GPUs, racks, and partnerships, challenging Nvidia. CEO Lisa Su unveiled Helios, a 72-GPU rack system featuring the MI455 GPU, with 12 compute/I/O chiplets, offering 20 PFLOPS FP8 compute, 432 GB HBM4 memory. The MI455 delivers 4x token throughput and up to 18x more tokens per dollar than its predecessor, achieving 10-15x better performance per watt. Major deployments include Anthropic acquiring 2 GW of Helios racks by H1 2027, and OpenAI deploying 6 GW of AMD hardware. AMD partnered with Cerebras for disaggregated inference, boosting fast token throughput by 5x. The company also detailed Venice CPUs, optimized for AI host nodes and agentic servers, with significant performance leads. The ROCm.AI software stack improved DeepSeek-R1 inference by 3.3x and training by 2.4x. AMD's roadmap includes the MI500 GPU next year, targeting over 2,000x higher inference throughput in four years.

Key takeaway

For AI Architects evaluating next-generation hardware, you should seriously consider AMD's new offerings. The Helios rack with MI455 GPUs provides compelling performance per dollar, and the ROCm.AI software stack significantly improves inference and training. Your teams can achieve 5x more throughput on fast tokens by combining AMD with Cerebras for disaggregated inference. This shift in the competitive landscape means you have more viable options for scaling AI infrastructure, potentially reducing vendor lock-in and optimizing cost-performance for diverse workloads.

Key insights

AMD is aggressively challenging Nvidia's AI hardware dominance with new GPUs, CPUs, and an open software ecosystem.

Principles

• Open ecosystems foster rapid AI development.
• Disaggregated inference optimizes throughput and latency.
• AI can automate kernel writing and system optimization.

Method

LLM inference can be disaggregated: GPUs handle prefill, dedicated accelerators manage decode for optimal throughput and latency.

In practice

• Deploy AMD Helios racks for high-throughput AI workloads.
• Combine AMD GPUs with Cerebras for low-latency token processing.
• Utilize ROCm.AI to accelerate GPU kernel development with AI agents.

Topics

• AMD Data Center GPUs
• AI Inference Hardware
• ROCm Software Stack
• Disaggregated AI
• AI Accelerators
• Large Language Models

Best for: CTO, VP of Engineering/Data, AI Engineer, AI Architect, Director of AI/ML, AI Hardware Engineer

Related on AIssential

• AMD makes case for integrated Helios AI stack with new GPUs, CPUs — AMD positions its integrated Helios stack as a comprehensive alternative to Nvidia for AI infrastructure, focusing on inference and agentic AI.
• Startup Boosts Scale-Up to 1000+ GPUs in a Single Domain — Disaggregated server design and software enable resilient, large-scale GPU inference clusters with flexible topologies.
• AMD Instinct™ GPUs MLPerf Inference v6.0 Submission — AMD's MI355X GPU and ROCm stack deliver competitive MLPerf Inference v6.0 results, especially for LLMs and multi-node setups.
• AMD Partners with Rival Cerebras on AI Server Rack — AMD and Cerebras partner for disaggregated inference, running AI across rival chips simultaneously.
• AMD unveils Helios AI rack to challenge Nvidia — AMD's Helios AI rack aims to disrupt Nvidia's AI hardware market dominance, driven by increasing demand for agentic AI compute.
• Intel: Our upcoming AI chip will be cheaper, run cooler than Nvidia, AMD options — Intel targets AI inference with a cost-effective, air-cooled GPU using LPDDR5 memory, aiming to undercut rivals.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Big Data & AI News - EE Times.