AMD INSTINCT™ MI355X

ULTIMATE AI & HPC PERFORMANCE
Built on the cutting-edge 4th Gen AMD CDNA™ architecture, AMD Instinct™ MI355X Platform features eight GPUs on a Universal Base Board (UBB) form factor. It features powerful and energy efficient cores with excellent performance per watt to drive the next era of AI and HPC innovation.

With exceptional efficiency and performance, the platform excels at high-speed inference, training massive AI models, and tackling the most complex computational challenges—from scientific simulations to large-scale data processing and advanced modeling.

DIRECT LIQUID COOLED DESIGN
The global adoption of AI technology demands ever greater amounts of processing capacity that can only be achieved by reaching for higher densities and more efficient approaches to power and cooling. While an air-cooled option is available, the direct liquid cooling option enables AMD Instinct MI355X GPUs to consume up to 1400W. Liquid cooling helps reduce GPU die temperatures, expanding the accelerator’s power envelope, facilitating higher clock speeds and delivered performance. Other component temperatures, for example the PCIe Gen 5 interfaces to the host system, are moderated by air-cooled heat sinks.

Supplied with bare GPU dies, server vendors install cold plates from one of two qualified vendors. This enables servers to use cooling technology they engineer. The use of cold plates instead of heat sinks for the individual GPUs enables the UBB to be integrated into standard and open compute servers, helping customers achieve higher densities in direct liquid cooled and air-cooled configurations.

INDUSTRY-STANDARD FORM FACTOR
The platform combines the power of eight accelerators on an industry-standard universal baseboard (UBB 2.0). The eight Open Compute Project (OCP) Accelerator Modules (OAMs) are connected with an AMD Infinity Fabric™ mesh that provides direct connectivity between each of the GPUs over 153.6 GB/s bidirectional links. Each MI355X connects with its peers through seven links, plus one PCIe® Gen 5 x16 connection per OAM device for upstream server and/or I/O connectivity. Each accelerator boasts 288 GB of HBM3E memory with 8 TB/s of memory bandwidth. This combines into a massive 2.3 TB of coherent, shared memory.

BREAKTHROUGH AI ACCELERATION
With the new expanded FP6 and FP4 datatype support, AMD Instinct MI355X GPUs maximize computational throughput, memory bandwidth utilization, and energy efficiency, enabling faster, more power-efficient AI inference compared to previousgeneration datatype support. Enhanced FP16 and FP8 processing, combined with added next-gen FP6 and FP4 capabilities, position the AMD Instinct MI350 Series products to deliver exceptional performance for advanced generative AI models—pushing the boundaries of AI acceleration.

SEAMLESS GPU UPGRADES FOR BETTER PERFORMANCE
The AMD GPU Operator simplifies deployment and management of AMD Instinct GPUs in Kubernetes clusters, helping enable effortless configuration of GPU-accelerated workloads, streamlining operations while accelerating time to market. Scaling software from prior-generation, air-cooled accelerators is simplified by the AMD ROCm™ platform. Day-0 support enables optimized models to be available upon product release.

ADVANCED SECURITY FOR AI & HPC DEPLOYMENTS
AMD Instinct MI355X GPUs deliver robust security to protect AI models, data, and system integrity. They help ensure only trusted firmware runs, verify hardware authenticity, enable secure multitenant GPU sharing, and encrypt high-speed GPU communication. These protections help enhance reliability, scalability, and data confidentiality, making them ideal for cloud AI, enterprise, and missioncritical workloads.

SCALABLE, FUTURE-PROOF NETWORKING
Ethernet-based AI and HPC networking enables massive hyperclass scalability, low TCO, and eliminates vendor lock-in, helping ensure open, flexible AI clusters.

OPTIMIZED SOFTWARE & FRAMEWORK INTEGRATION
The foundation of AMD accelerated computing, AMD ROCm software empowers AI developers to fully leverage AMD Instinct GPUs including the latest Instinct MI350 Series for both inference and training. With Day-0 support for major AI frameworks like PyTorch, TensorFlow, JAX, and ONNX Runtime, the latest advancements in ROCm helps simplify AI model migration and deployment, optimizing hardware efficiency while minimizing code changes. Through strategic collaborations with AI leaders such as OpenAI, PyTorch, Hugging Face, Databricks and more, ROCm delivers high-performance, out-of-the-box AI solutions, enabling smooth integration, and robust partner support. The ROCm 7.0 release delivers significant advancements in AI training and infererence. Mixed-precision training is optimized through support for FP16, FP8, BF16, FP6, FP4 datatypes, helping accelerate training while maintaining high accuracy for AI models. Fast data loading and augmentation pipelines can reduce CPU-GPU bottlenecks, helping ensure high-throughput data preprocessing for large datasets. Once models are trained, AI inference can be optimized with low-latency kernels to help enhance transformer-based model execution, acceleratingattention mechanisms for fast response times. With enhanced quantization support, ROCm software is designed to improve FP8, FP6, and FP4 inference, enabling efficient low-bit AI processing that reduces power consumption while maintaining accuracy, helping enable real-time processing, ultra-low latency, and increased power efficiency.

TRUSTED BY AI LEADERS
AI giants including Meta have chosen AMD Instinct™ GPUs for their own large-scale AI deployments, powering models like Llama 405B and GPT. With adoption across many leading and emerging cloud service providers, original equipment manufacturers, and original design manufacturers, AMD is driving next-gen AI at scale.

LEARN MORE
The AMD Instinct MI355X Platform is available through AMD solution partners. Standard form factors facilitate adoption into enterprise servers so that you can use the same power in your data center, or in the cloud from the leading superscalars. Learn more at AMD.com/INSTINCT.