Particle.news
Download on the App Store
4 ARTICLES
·
6d ago

AMD Unveils Instinct MI350X and MI355X AI Accelerators

It demonstrates AMD’s strategy to rival Nvidia by marrying enhanced inference throughput with an open-standard rack ecosystem.

Image
Image
Image

Overview

  • The MI350X and MI355X leverage TSMC’s N3P process and a redesigned chiplet package to deliver up to 4.2 times faster AI inference than their MI300X predecessors while improving power efficiency.
  • The MI355X variant is engineered for high-density racks with direct liquid cooling and can accommodate up to 128 GPUs at up to 1.4 kW per unit.
  • AMD revealed plans for its Helios rack-scale architecture for next year’s MI400 accelerators, designed for both scale-up and scale-out configurations using open UBB standards.
  • The forthcoming MI400 GPUs are projected to double FP4 throughput over the MI355X and integrate 432 GB of HBM4 memory with enhanced per-GPU bandwidth.
  • Beyond 2025, AMD’s roadmap extends to 2027 with the Instinct MI500 series and Epyc Verano processors aimed at expanding its market share against Nvidia.