Particle.news
Download on the App Store
9 ARTICLES
·
last mo.

AMD Launches Instinct MI350 Series AI GPUs With 4× AI Compute Gains

Built on CDNA4 with a 3nm process node, the MI350 series delivers up to 35× faster inference versus the MI300X

Lisa Su, CEO of AMD, attends the Artificial Intelligence (AI) Action Summit at the Grand Palais in Paris, France, February 10, 2025. REUTERS/Benoit Tessier/File Photo
blank
Image
Image

Overview

  • The MI350X and MI355X pack 185 billion transistors and up to 288 GB of HBM3E memory on TSMC’s 3nm node for enhanced throughput and efficiency.
  • AMD reports a fourfold boost in AI compute performance and a 35× leap in inference speed over the MI300X, enabled by new FP4 and FP6 data-type support and up to 20 PFLOPs of low-precision compute.
  • The MI350X is optimized for air-cooled deployments with lower board power, while the liquid-cooled MI355X peaks at 1,400 W and offers 8 TB/s of memory bandwidth.
  • Benchmarks from AMD show up to 1.3× faster inference and up to 1.1× training gains against Nvidia’s B200 and GB200 accelerators in key AI workloads.
  • The MI350 series will ship through OEM partners starting in Q3 2025 and paves the way for AMD’s next-generation MI400 GPUs slated for a 2026 debut.