At the Advancing AI 2025 event held in San Jose, California, AMD announced the official launch of its CDNA 4 architecture design and up to 288GB of HBM3e high-bandwidth memory.Instinct MI350 series, a new generation of high-performance computing platform accelerators, boasting a 35x improvement in AI inference performance (compared to the Instinct MI300 series). AMD also reiterated that the Instinct MI400 series accelerator will be launched in 2026 and will feature the next-generation AI architecture codenamed "Helios."
Competing with NVIDIA in the demand for AI accelerated inference
Instinct MI350系列區分Instinct MI355X、Instinct MI350X設計，兩者均採用CDNA 4架構設計，並且都搭載由美光、三星提供的288GB HBM3e高頻寬記憶體，支援每秒8TB的資料傳輸速率，差異性則各有不同，例如Instinct MI355X在FP64峰值算力可達79TFLOPS，FP4算力表現可達20PFLOPS，而Instinct MI350X則分別可達72TFLOPS、18.4PFLOPS。
In addition, the thermal design power consumption of Instinct MI355X is up to 1400W, while that of Instinct MI350X is up to 1000W.
Both adopt OAM module design and support water cooling or air cooling architecture. Up to 64 GPUs can be configured in air cooling architecture, and up to 128 GPUs in water cooling architecture. Compared with NVIDIA's GB200 NV72, it can be configured with more GPUs for computing acceleration, and the memory capacity is also larger than GB200 NV72. The peak computing power performance also surpasses GB4 NV8 in FP200 and FP72 computing.
However, compared to NVIDIA's GB200 NV72, which uses the Arm-based Grace CPU, AMD's computing architecture uses the fifth-generation EPYC server processor code-named "Turin," an x86 architecture. It is also paired with the Pensando network interface controller code-named "Pollara," which can expand transmission bandwidth. It is also compatible with the Open Computing Platform (OCP) open computing architecture and is expected to be used in Oracle's Oracle Cloud Infrastructure (OCI) and other hyperscale cloud provider service platforms.
The ROCm 7.0 open-source AI software platform enhances AI inference capabilities and builds a cloud-hosted environment for developers.
In conjunction with the newly launched Instinct MI350 series accelerators, AMD also announced the release of the ROCm 7.0 open source artificial intelligence software platform, which is claimed to increase artificial intelligence inference capabilities by up to 3.5 times. It also supports the FP4 computing format and the FA3 (Flash Attention 3) algorithm, enabling open source frameworks such as vLLM virtual large language models and SGLang to run more quickly.
AMD also announced the launch of a cloud-hosted developer environment. This environment, which allows developers to more easily build various application services using the ROCm software stack and AMD GPUs, includes the ready-to-use Docker container platform, ready-to-use vVLLM, SGLang, and HAO AI Lab.
Instinct MI400 series accelerator, codenamed "Helios" rack configuration design
AMD has once again announced that it will launch the Instinct MI2026 series accelerators in 400, and has proposed a rack configuration design code-named "Helios", which will integrate the Instinct MI400 series accelerators with AMD's next-generation EPYC server processor code-named "Venice" and the Pensando network interface controller code-named "Vulcano" based on the Zen 6 architecture. This will compete with NVIDIA's future artificial intelligence rack design.
Many companies announced their cooperation with AMD
At the event, companies including Meta, OpenAI, Microsoft, and xAI all expressed their commitment to using AMD solutions to advance AI applications:
• Meta detailed its extensive deployment of the Instinct MI300X for Llama 3 and Llama 4 inference. Meta is excited about the MI350 and its computing power, performance per total cost of ownership (TCO), and next-generation memory. Meta will continue to work closely with AMD to advance the AI roadmap, including plans for the Instinct MI400 series platform.
• OpenAI CEO Sam Altman explained the importance of comprehensive optimization of hardware, software and algorithms, and emphasized the close collaboration between OpenAI and AMD on AI infrastructure, including the use of MI300X to develop Azure Research and GPT models, and deep design collaboration on the MI400 series platform.
• Oracle Cloud Infrastructure (OCI) is one of the first industry leaders to adopt AMD's open rack-scale AI infrastructure, powered by the AMD Instinct MI355X GPU. Leveraging AMD CPUs and GPUs to deliver balanced, scalable performance for AI clusters, OCI announced the availability of Zettascale AI clusters accelerated by the latest AMD Instinct processors, featuring up to 131,072 MI355X GPUs, enabling customers to achieve large-scale AI builds, training, and inference.
• HUMAIN shared its landmark agreement with AMD to build an open, scalable, flexible, and cost-effective AI infrastructure using a comprehensive computing platform that only AMD can provide.
•Microsoft announced that the Instinct MI300X is now available for productivity for proprietary and open source models on Azure.
•Cohere noted that its high-performance, scalable Command model has been deployed on the Instinct MI300X, powering enterprise-grade LLM inference with high throughput, efficiency, and data privacy.
• Red Hat described how its expanded collaboration with AMD enables a production-ready AI environment, delivering powerful, efficient AI processing in hybrid cloud environments through AMD Instinct GPUs on Red Hat OpenShift AI.
•Astera Labs highlighted how the open UALink ecosystem accelerates innovation and brings greater value to customers, and shared plans to provide a comprehensive UALink product portfolio to support the next generation of AI infrastructure.
•As a member of the UALink Alliance, Marvell and AMD are collaborating to develop open interconnect technologies that bring ultimate flexibility to AI infrastructure.
