Skip to main content

AMD Instinct MI350X / MI355X (CDNA 4)

Overview

The AMD Instinct MI350 series, launched on June 12, 2025, is AMD's latest flagship AI accelerator, based on the CDNA 4 architecture. It features 288GB HBM3e memory, 8 TB/s bandwidth, and up to 2.5 PFLOPS FP16 compute. The MI350 introduces native MXFP4 and MXFP6 data type support, targeting NVIDIA Blackwell.

Core Specifications

ItemMI350XMI355X
ArchitectureCDNA 4CDNA 4
Memory288 GB HBM3e288 GB HBM3e
Bandwidth8 TB/s8 TB/s
LLC256 MB256 MB
FP6472.1 TFLOPS78.6 TFLOPS
FP16 Matrix2.3 PFLOPS2.5 PFLOPS
MXFP84.6 PFLOPS5.0 PFLOPS
MXFP69.2 PFLOPS10.1 PFLOPS
MXFP49.2 PFLOPS10.1 PFLOPS
TDP750 W1,400 W (liquid cooling)
CoolingAirLiquid (DLC)
PCIeGen 5Gen 5

Vendor Information

ItemDetail
ManufacturerAMD, Inc.
Websitehttps://www.amd.com
Product Pagehttps://www.amd.com/en/products/accelerators/instinct/mi350.html
ReleaseJune 12, 2025
ROCm 7Released concurrently

MI350 Platform (UBB 2.0)

PlatformConfiguration
8× MI350X2.3 TB HBM3E, 64 TB/s aggregate bandwidth, 18.5 PFLOPS FP16
8× MI355X2.3 TB HBM3E, 64 TB/s aggregate bandwidth, 20.1 PFLOPS FP16
Hyperscale128× liquid-cooled MI355X (16 UBB 8-card trays)

MI350 vs Blackwell Comparison

MetricMI355XB200GB200 (186GB)
Memory288 GB192 GB186 GB
Bandwidth8 TB/s8 TB/s8 TB/s
FP8 Compute10.1 PFLOPS (MXFP6)9 PFLOPS9 PFLOPS
FP4 Support✓ MXFP4✓ Native✓ Native

Key difference: MI355X has more memory (+50%), while B200 has higher FP4 compute (18 PFLOPS vs MI355X 10.1 PFLOPS).

Key Features

  • Native MXFP4 / MXFP6 support: Micro-scaled floating-point formats
  • 4× previous-gen AI compute (AMD official)
  • 35× previous-gen inference performance
  • CDNA 4 architecture (N3P process XCD)
  • ROCm 7: New pip install rocm simplified installation

Use Cases

  • Trillion-parameter LLM training and inference
  • Long-context inference
  • MoE models (Mixtral, Llama 4)
  • HPC and AI mixed workloads