AMD launches Instinct MI350 GPUs to accelerate generative AI models

AMD's new GPUs for AI workloads

During the Advancing AI 2025 event, AMD introduced the Instinct MI350 series of graphics processors.

This new generation, based on the CDNA 4 architecture and a 3-nanometer process node, consists of the MI350X and MI355X models.

Both models include 288GB of HBM3E memory and 8TB/s bandwidth, enabling models with up to 520 billion parameters to be processed on a single GPU. The MI355X achieves up to 20 PFLOPs in FP4 and FP6 precision, with a power consumption of up to 1400W.

Compared to the MI300X model, the new series delivers up to 4 times the performance on inference tasks and up to 3.5 times the performance on training models such as Llama 3.1 and DeepSeek R1.

This is made possible by improvements in computing units, increased memory bandwidth, and new data types such as FP4, FP6, and FP8.

Architecture optimized for efficiency and scalability

The MI350s are built with a chiplet architecture consisting of a 3nm compute die (XCD) and two 6nm input/output dies (IOD), connected using hybrid 3D packaging technology. They also integrate the fourth generation of Infinity Fabric and 256 compute units.

The design supports GPU partitioning for multiple loads and is ready to run on infrastructures with up to eight concurrent instances.