Built for AI Factories

AMD Instinct™ MI455X

Power your AI workloads with MI455X in Helios™ rack-scale architecture, combining next-generation HBM4, massive memory bandwidth, and ROCm-optimized AMD infrastructure for frontier-scale training and inference.

AMD Instinct MI355X GPU

MI455X on TensorWave

Experience true rack-scale dominance with AMD's groundbreaking Helios architecture. Integrating up to 72 MI400-series GPUs (such as the MI455X) into a single rack, providing over 31 TB of aggregate HBM4 memory and high-bandwidth connectivity through optimized direct liquid cooling, UALink interconnects, and an Ethernet-forward approach. This dense, cohesive design supports large deployments more efficiently than traditional multi-rack setups.

Bullet Point Icon

Unmatched High-Memory Power

Bullet Point Icon

Helios Rack-Scale Design

Bullet Point Icon

Built for Production at Scale

AMD Instinct™ MI400 Series announcement

"MI455 is the most advanced chip we've ever built... It includes 320 billion transistors, 70% more than MI355. It includes 12 two-nanometer and three-nanometer compute and IO chiplets, and 432 gigabytes of ultra-fast HBM4."

Dr. Lisa Su

CEO, AMD

Quote Image

AMD Instinct MI455X

Server specifications can be customized to your needs.

DIRECT LIQUID COOLING

Pricing
TBD
Expected Memory: 432 GB HBM4, 19.6 TB/s bandwidth
Expected Performance: ~40 PFLOPS FP4 (dense), ~20 PFLOPS FP8
Expected Architecture: 12x HBM4 stacks, 320B transistors, 2nm/3nm chiplets, 3.5D packaging

Unlock significantly more memory and scale your AI workloads.

Related Blog Posts