- AMD launches the Instinct MI350 series, featuring the MI350X and MI355X models, with up to 35x improvement in inference over the previous generation.
- The new GPUs are manufactured in 3nm, feature 185.000 billion transistors and 288GB of HBM3E memory, and feature outstanding performance and power efficiency.
- Supermicro and major technology partners such as Meta, Microsoft, OpenAI and Oracle already adopt these solutions to train and scale advanced models of IA.
- The MI350-based infrastructure and ROCm 7 software provide flexibility, scalability, and dramatic reductions in the cost and energy consumption of AI data centers.
AMD has taken a decisive step in its strategy to compete in the field of Artificial Intelligence with the Launch of the AMD Instinct MI350 accelerator family. This new series, presented at the Advancing AI 2025 event, comes with the promise of becoming one of the benchmarks in performance, efficiency and scalability for computing tasks. Generative AI and high-performance computing. Thanks to a combination of hardware Advanced and open software ecosystem, AMD strengthens its position against other major manufacturers and manages to capture the attention of technological giants such as Meta, Microsoft, OpenAI and Oracle.
The MI350 series represents a significant generational leap forward, improving performance by up to 4 times over the previous family and achieving up to 35 times greater efficiency in inference processes.This radical leap is intended to meet the growing demand for power in data centers and AI applications, which require processing and training increasingly complex models.
Technical news of the AMD Instinct MI350 series
El The heart of the series is made up of the MI350X and MI355X models, both manufactured with the 3 nanometer process of TSMC. They integrate 185 billion transistors, enabling remarkable density and energy efficiency. Each GPU is equipped with 288 GB HBM3E memory (supported by Micron and Samsung), capable of achieving a Bandwidth up to 8 TB/s.
In terms of architecture, the new cards base their design on 256 computing units and 16.384 cores, distributed across eight zones with independent XCDs. They also leverage fourth-generation Infinity Fabric interconnect technology and a 256MB Infinity Cache, optimizing data transfer for AI-intensive workloads.
Another differentiating aspect is the compatibility with advanced data types such as FP4 and FP6, ideal for large-scale model inference. The series covers a variety of cooling needs, as it is available in air- and liquid-cooled versions, allowing for deployment in different rack configurations. A single rack can accommodate up to 64 air-cooled GPUs or up to 128 GPUs if direct liquid cooling is chosen, achieving impressive 2,6 exaFLOPS of FP4/FP6 performance in the set.
Comparisons with the competition and energy efficiency
AMD has presented data that places the MI355X ahead of B200 cards Nvidia in performance, being up to 2,2 times faster in certain scenarios. In inference tests with AI models such as Llama 3.1 405B, the company claims that the MI355X increases its performance by 35 times compared to the previous MI300 series and outperforms the competition in both calculation speed and memory capacity, with a 40% more tokens generated per dollar invested.
Energy efficiency is also a strong point of this generation. According to AMD, the MI350 series has achieved up to 38-fold improvements in power efficiency compared to the base generation five years ago, exceeding its planned targets. It aims to increase this efficiency by 20-fold by 2030 at the rack level. This could allow, in a few years, a model that currently requires hundreds of racks for training to be managed with a fraction of those resources.
Infrastructure, ecosystem and collaborations with major partners
Integration into data center platforms is another of AMD's pillars. Manufacturers such as Supermicro now offers optimized server solutions with the MI350 GPUs, available in air- or liquid-cooled versions and ready to scale up to hundreds of GPUs per rack. These platforms utilize AMD EPYC CPUs and Pollara interconnect technology to maximize performance, scalability, and overall system efficiency, facilitating deployment in the cloud and enterprises requiring large-scale AI.
Major technology companies have already opted for AMD technology for various purposes. Meta has implemented the previous generation (MI300X) for inference of its Llama models and plans to move to the MI350 series. as soon as it is available, highlighting computing power and memory. AI performance on Windows systems It is also boosted by advances like these. Microsoft Azure uses AMD accelerators to run both proprietary and open AI models in production, and Oracle Cloud Infrastructure will deploy massive clusters powered by MI355X. to achieve zettascale. For their part, OpenAI and Cohere have highlighted the efficiency and performance that AMD solutions provide.
In the software field, the renewed version of ROCm 7 Boosts AI performance and facilitates compatibility with current major frameworks and models such as Llama or DeepSeekAdditionally, the company has launched AMD Developer Cloud, a cloud platform optimized for AI developers and research.
Preparing for the future: roadmap and new generations
AMD is not stopping there and has already advanced the development of the next generation of accelerators, the Instinct MI400 series, based on HBM4 memory and an even more ambitious architecture. This line will be available alongside "Helios" infrastructure, EPYC Venice (Zen 6) processors, and new Vulcan Thinking network cards, with the goal of increasing the performance of inference systems tenfold.
With all these new developments, AMD seems determined to democratize access to artificial intelligence and position itself as one of the main players in the open AI ecosystem, both in terms of hardware and software, and support for the technological community.
Instinct MI350 accelerators will be available to the firm's partners during the third quarter of 2025, and their adoption in data centers and cloud services is expected to mark a turning point in the way advanced artificial intelligence challenges are addressed. The combination of raw performance, efficiency, and infrastructure flexibility, along with the support of major companies in the sector, indicates that the impact of this technology will be considerable in the coming years.
Passionate writer about the world of bytes and technology in general. I love sharing my knowledge through writing, and that's what I'll do on this blog, show you all the most interesting things about gadgets, software, hardware, tech trends, and more. My goal is to help you navigate the digital world in a simple and entertaining way.