On January 6th, during the keynote speech at CES 2026, AMD not only presented the technical intricacies of its next-generation Instinct MI500 series AI accelerators but also set forth an audacious objective: to amplify AI performance by a factor of 1000 within a mere four years.

The forthcoming MI500 series, slated for release in 2027, will be crafted using the cutting-edge 2nm process technology and will boast a revolutionary CDNA 6 architecture. This will be complemented by next-generation HBM4E memory, delivering bandwidth that surpasses the 19.6 TB/s offered by the current MI400 series.

AMD's CEO, Lisa Su, painted a vivid picture in her address: since the advent of ChatGPT, the number of active AI users has skyrocketed from 1 million to a staggering 1 billion, a feat that took the internet decades to accomplish.

She foresees that by 2030, the number of active AI users will soar to 5 billion. To accommodate this exponential growth, "we must augment the world's computing capacity by a factor of 100," she asserted.

AMD's answer to this challenge is the MI500 accelerator. Set to debut in 2027, this product represents another monumental leap for AMD in the realm of AI computing.

The MI500 series will showcase a brand-new CDNA 6 architecture, an evolution of AMD's architecture meticulously crafted to expedite computationally intensive AI and high-performance computing tasks.

When juxtaposed with the CDNA 5 architecture utilized in the upcoming MI400 series, the CDNA 6 architecture is anticipated to yield even more substantial performance enhancements.

This accelerator will be outfitted with HBM4E memory, offering speeds and bandwidth that eclipse the 19.6 TB/s provided by the MI400 accelerator, which is based on HBM4.

During the CES 2026 demonstration, AMD displayed a compelling performance growth chart, clearly delineating the performance evolution trajectory from the MI300X to the MI500 series.

Adjacent to the chart, a bold target of enhancing AI performance by 1000 times within four years was prominently featured. This commitment is not directed at a single product generation but rather at the cumulative impact of the entire product line upgrade chain.

From the MI300X in 2023 to the MI325X in 2024, followed by the MI350 series in 2025, the MI400 series in 2026, and culminating with the MI500 series in 2027, AMD is establishing a predictable cadence of platform evolution.

Hardware advancements constitute just one facet of AMD's strategy. The software ecosystem is equally vital to its competitiveness.

AMD's ROCm software stack has witnessed significant strides over the past year. In 2025, AMD extended ROCm platform support to Windows systems and encompassed a broader range of Linux distributions.

This expansion resulted in a tenfold surge in ROCm downloads year-over-year. The launch of ROCm 7 further achieved performance improvements of up to 3.8 times in inference and 3 times in training.

AMD also introduced the AMD Software: Adrenalin Edition AI Bundle, a novel optional feature aimed at simplifying and expediting the setup of local AI environments.

At CES 2026, AMD showcased not merely a solitary product but a comprehensive AI product portfolio spanning from data centers to the edge.

AMD unveiled its Helios rack-level platform for the first time, a blueprint tailored for yotta-scale AI infrastructure.

Constructed on AMD Instinct MI455X GPUs and AMD EPYC 'Venice' CPUs, this platform is specifically engineered for advanced AI workloads, delivering up to 3 AI exaflops of performance per rack.

AMD also broadened its enterprise-grade AI accelerator lineup, introducing the AMD Instinct MI440X GPU designed for local enterprise AI deployments.

On the consumer front, AMD launched the new-generation Ryzen AI 400 series platform, offering 60 TOPS of NPU computing power and full support for the AMD ROCm platform, facilitating seamless AI expansion from the cloud to the client.

Notably, through its Ryzen AI Max+ series processors, AMD enables Windows AI PCs equipped with 128GB of unified memory to run large language models with up to 128 billion parameters locally.

This breakthrough signifies the first instance where a consumer-grade processor can execute models of such magnitude, bringing unprecedented flexibility to local AI deployments.

AMD's stated goal of boosting AI performance by 1000 times in four years not only underscores its confidence in hardware technology but also its holistic approach to software ecosystems and platform strategies.

AMD appears to be maintaining the CDNA-centric product line naming system for its Instinct GPUs, rather than switching to UDNA.

This decision preserves brand consistency while also showcasing AMD's long-term strategic vision in the AI accelerator market.

References:

https://www.ithome.com/0/910/798.htm

https://www.ithome.com/0/910/785.htm

https://ir.amd.com/news-events/press-releases/detail/1272/amd-and-its-partners-share-their-vision-for-ai-everywhere-for-everyone-at-ces-2026