Particle.news
Download on the App Store

AMD Launches Instinct MI350P PCIe AI Accelerator With 144GB HBM3E

The drop-in card targets on‑prem AI inference in standard servers to challenge Nvidia's H200 NVL.

Overview

  • AMD introduced the Instinct MI350P on Thursday as a dual‑slot PCIe accelerator built for easy installation in existing air‑cooled servers.
  • The CDNA 4 card has 128 compute units (8,192 cores) on a 3nm compute die with a 6nm IO die, 144 GB of HBM3E at 4 TB/s, and a 600W board power.
  • AMD cites up to 4.6 petaFLOPS peak in MXFP4 with native MXFP6/MXFP4 support and sparsity features to speed large language model inference.
  • Positioned against Nvidia’s H200 NVL PCIe card, coverage notes roughly 20% higher FP64 and about 40% higher FP16 and FP8 theoretical compute, while Nvidia has not announced an HBM Blackwell PCIe rival.
  • AMD says the card is available through partners with support for its ROCm software stack, though real‑world uptake will depend on tooling maturity and buyer preferences.