AMD INSTINCT™ MI350 SERIES GPUs: 
FUTURE-READY YOUR DATA CENTER 
FOR AI ADVANCES

AMD Instinct MI350 Series GPUs set a new standard for generative AI and HPC applications in data centers. Built on the cutting-edge 4th Gen AMD CDNA™ architecture, these GPUs deliver exceptional efficiency and performance for training massive AI models, high-speed inference and complex HPC workloads. At the same time, the AMD ROCm™ platform, the industry’s premier open-source alternative for AI and HPC, allows AMD Instinct Series GPUs to easily integrate with multiple operating environments and platforms, enabling effortless AI and HPC deployment with minimal code changes.

LEADERSHIP PERFORMANCE FOR AI AND HPC
Discover the possibilities of accelerated performance at scale.

AMD introduces the new MI350 Series GPUs, consisting of the AMD Instinct MI350X GPU, delivering leading performance for AI inference and training at scale, as well as the AMD Instinct MI355X GPU, designed for high-density, efficient and scalable infrastructure. With enhanced memory capacity, robust datatype support and multiple cooling options, the MI350 Series GPUs handle diverse AI workloads efficiently, from large-scale inferencing to generative AI model training.

SEAMLESS SCALABILITY AND DEPLOYMENT
Ease the configuration of GPU-accelerated workloads such as machine learning and generative AI.

Air cooled AMD Instinct MI350 Series GPUs integrate seamlessly with prior generation AMD Instinct MI300 Series platforms (MI300X and MI325X) and competitive infrastructures, offering cost-effective performance upgrades for evolving infrastructure demands. They work with the AMD GPU Operator, which simplifies the deployment and management of AMD Instinct GPU accelerators within Kubernetes clusters.

TRUSTED BY AI LEADERS
Join the community of CSPs, OEMs and ODMs driving next-gen AI at scale.

AMD Instinct Series GPUs power AI, driving adoption across leading cloud service providers (CSPs) and original equipment manufacturer (OEMs) partners. AMD delivers scalable, efficient AI solutions, empowering customers with greater choice and a competitive edge in AI innovation. Their adoption underscores their proven performance, efficiency and scalability and reinforces AMD as a trusted supplier for next-generation AI.

PROVEN, OPEN, LIMITLESS AI SOFTWARE
Enable effortless AI and HPC deployment with minimal code changes and maximum flexibility.

AMD facilitates the adoption and use of multiple acceleration platforms and cross-platform HPC and AI development with the ROCm 7.0 open-source software ecosystem and programming toolset. A collection of drivers, development tools and APIs, it extracts the best performance from AMD Instinct GPUs while maintaining compatibility with industry software frameworks.

BUILT-IN SECURITY FOR AI AND HPC DEPLOYMENTS
Get advanced security to protect AI models, data and system integrity.

AMD Instinct MI350 Series GPUs deliver robust security features to protect AI models, data and system integrity. They help ensure only trusted firmware runs, verify hardware authenticity and encrypt high-speed GPU communication. These protections make them ideal for cloud AI, enterprise and mission-critical workloads.

TECHNICAL DEEP DIVE

#1 DISCOVER THE POSSIBILITIES OF ACCELERATED PERFORMANCE AT SCALE

• AMD Instinct MI350 Series GPUs feature powerful and energy-efficient cores, maximizing performance per watt to drive the next era of AI and HPC innovation. With newly expanded support for FP6 and FP4 datatypes, a massive 288GB of HBM3E memory and up to 8TB/s bandwidth, they maximize computational throughput, memory bandwidth utilization and energy.

• AMD Instinct MI350X and MI355X 8xGPU platforms offer up to 36.8 and 40.3 PFLOPs bfloat16 performance, respectively, with structured sparsity, and with new expanded support for new datatypes, they offer up to 148 and 161.1 PFLOPs FP6/FP4 matrix performance, respectively, with structured sparsity.

• AMD Instinct MI350X and MI355X 8xGPU platforms offer up to 2.05x and 2.2x the peak theoretical FP6 matrix PFLOPS performance, respectively, vs. the NVIDIA HGX B200 GPU.

• The AMD Instinct MI355X GPU offers up to 2x the peak theoretical FP6 matrix PFLOPS performance with structured sparsity vs. the NVIDIA GB200 GPU.

• AMD Instinct MI350X and MI355X GPUs offer up to 77% and 93% higher peak theoretical FP8 matrix performance, respectively, with structured sparsity than the previous-generation AMD Instinct MI300X GPU.

• AMD Instinct MI350X and MI355X GPUs with FP4 matrix datatype are projected to deliver up to 7x and 7.7x the performance, respectively, on AI and training workloads compared to the AMD Instinct MI300X GPU using FP16 matrix.

• AMD Instinct MI350 Series GPUs offer up to 2.04x the memory capacity of the NVIDIA H200 GPU and up to 60% more memory capacity than the NVIDIA HGX B200 GPU.

• AMD Instinct MI350 Series GPUs offer up to 50% more memory capacity than previous generation AMD Instinct MI300X GPUs.

• AMD data center solutions support AI and HPC workloads at tremendous scale with enhanced efficiency. AMD powers 5 of the top 10 supercomputer systems on the Top500 list and more than half of the top 20 supercomputer systems on the Green500 list.


#2 EASE THE CONFIGURATION OF GPU-ACCELERATED WORKLOADS SUCH AS MACHINE LEARNING AND GENERATIVE AI

• AMD Instinct MI350 Series GPUs optimize efficiency, bandwidth and energy use for fast, power-efficient AI inference and training.

• Enhanced processing with next-gen FP6 and FP4 capabilities enables AMD Instinct MI350 Series GPUs to deliver exceptional performance for generative AI models, pushing the boundaries of AI acceleration.

• Support for matrix sparsity further helps optimize AI training and inference, helping scale AI solutions across data centers.


#3 JOIN THE COMMUNITY OF CSPs, OEMs AND ODMs DRIVING NEXT-GEN AI AT SCALE

• With support for third-party partners such as Clarifai, ClearML, Fireworks, Flex.ai, Lamini, Rapt.ai, RHEL, OpenShift and UbiOps, AMD provides customers with flexibility, choice and best-in-class tools to accelerate AI innovation.


#4 SMOOTH YOUR AI AND HPC DEPLOYMENTS WITH MINIMAL CODE CHANGES AND MAXIMUM FLEXIBILITY

• ROCm 7.0 lets you develop applications in an open-source system so that the software is portable, allowing movement among GPUs from different vendors or among different inter-GPU connectivity architectures—all in a device-independent manner, unlike NVIDIA’s proprietary hardware solutions.

• The latest ROCm enhancements optimize AI inference, training and framework compatibility, delivering high throughput and ultra-low latency for workloads such as Natural Language Processing (NLP), computer vision and beyond. ROCm is well-suited to GPU-accelerated HPC, AI, scientific computing and computer-aided design (CAD).

• AMD Instinct MI350 Series GPUs easily integrate with JAX, Kokkos, ONNX, PyTorch, Raja, Runtime, TensorFlow, Triton, vLLM and more, enabling effortless AI and HPC deployment with minimal code changes.

• With Day 0 support, ROCm software helps optimize AI workloads from Hugging Face, Meta, OpenAI and others, delivering high throughput, low latency and peak efficiency for NLP, computer vision and generative AI inference—accelerating AI innovation at scale


#5 GET ADVANCED SECURITY TO PROTECT AI MODELS, DATA AND SYSTEM INTEGRITY

• AMD Instinct MI350 Series GPUs offer Device Secure Boot and Secure Update and Recovery to help ensure only trusted firmware runs, while Platform-Level DICE Identity and Attestation verify GPU authenticity to prevent unauthorized access.

• Infinity Fabric™ Link Security helps ensure high-speed GPU-to-GPU communication that’s safe.

1 Top500 list, November 2024, https://top500.org/lists/top500/2024/11/.

2 Green500 list, November 2024, https://top500.org/lists/green500/list/2024/11/.

©2025 Advanced Micro Devices, Inc. All rights reserved. AMD, the AMD arrow, AMD Instinct, Infinity Fabric, AMD Infinity Platform, AMD CDNA, ROCm and combinations thereof, are trademarks of Advanced Micro Devices, Inc. NVIDIA is a trademark of NVIDIA Corporation in the U.S. and/or other countries. The OpenMP name and the OpenMP logo are registered trademarks of the OpenMP Architecture Review Board. Other product names used in this publication are for identification purposes only and may be trademarks of their respective companies.