AMD Instinct

Overview

AMD Instinct is AMD’s line of GPU accelerators for datacenter AI training, inference, and HPC workloads. Built on successive generations of AMD’s CDNA architecture, the line runs from the MI200 series through MI300 (shipping 2023–2024), MI350 (shipping 2H 2025), and the forthcoming MI400 generation anchored by the Helios rack-scale system (target H2 2026). AMD competes directly with NVIDIA’s Hopper/Blackwell line and positions on memory capacity advantages and open-standards interconnect.

Entries

Entries

  • AMD Instinct — Overview — Architecture generations, competitive positioning, and market context for AMD's Instinct GPU accelerator line.
  • AMD Instinct MI300 Series — MI300X and MI300A: AMD's CDNA 3 generation GPU accelerators featuring 192 GB HBM3 memory; the fastest-ramping product in AMD history.
  • AMD Instinct MI350 Series — MI350X and MI355X: AMD's CDNA 4 generation GPU accelerators with 288 GB HBM3e and MXFP4 support, shipping 2H 2025.
  • AMD Instinct MI400 Series & Helios Rack — AMD's next-generation CDNA 5 GPU accelerators (MI430X, MI440X, MI455X) and Helios rack-scale AI system; 72 GPUs, 31 TB HBM4, 2.9 exaFLOPS FP4; targeting H2 2026.
  • AMD Pensando Pollara 400 AI NIC — AMD's 400G AI network interface card for scale-out AI cluster networking, UltraEthernet Consortium compliant, deployed in Instinct MI350 and MI400 infrastructure.
  • ROCm — AMD GPU Software Platform — AMD's open-source GPU software stack for AI and HPC workloads, competing with CUDA. Includes HIP (CUDA compatibility layer), ROCm runtime, and AI libraries.