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

AMD Unveils MI400X GPUs and Helios AI Rack to Rival Nvidia's Vera Rubin Systems

Pooling 72 MI400 accelerators via Ultra Accelerator Link into one rack, Helios delivers higher compute density with 50 percent more memory bandwidth than Nvidia's Vera Rubin NVL144.

AMD
blank
Image

Overview

  • The Instinct MI400X GPU boasts up to 10× the compute power of the MI300X and doubles FP4 performance over the MI355X, featuring 432 GB of HBM4 and 19.6 TBps of memory bandwidth.
  • Helios integrates 72 MI400-series GPUs linked by UALink, paired with EPYC “Venice” CPUs and Pensando Vulcano NICs, to reach 2.9 exaFLOPS of FP4 inference and 1.4 exaFLOPS of FP8 training performance.
  • AMD claims the new rack offers 50 percent more memory capacity and bandwidth—31 TBps and 1.4 PBps—compared with Nvidia’s upcoming NVL144 Vera Rubin platform.
  • The company has moved to a yearly cadence for its AI hardware roadmap and plans to ship EPYC “Verano” CPUs, Instinct MI500-series GPUs, and a denser rack-scale system in 2027.
  • AMD expects to leverage TSMC’s A16 process node for its 2027 chips, gaining backside power delivery to boost efficiency and performance density in future AI deployments.