AMD Unveils Instinct MI455X with CDNA 5, HBM4, and Helios Rack-Scale AI Systems

AMD has officially launched the Instinct MI455X accelerator, integrating the new CDNA 5 architecture on TSMC’s 2nm/3nm process technology. This release delivers significant compute density and memory bandwidth uplift, explicitly targeting large-scale AI and HPC deployments.

CDNA 5 Architecture and Compute Density

The MI455X leverages the CDNA 5 architecture, fabricated using TSMC’s advanced 2nm process for its compute dies (XCDs) and 3nm for its I/O die, fabric, and cache dies. This new design features 320 billion transistors and delivers a 4.2x increase in FP8/FP16 matrix math throughput compared to its predecessor, the MI300X. The accelerator utilizes an 8-die multi-chip module (MCM) interconnect for enhanced parallelism, with 256 Work Group Processors (WGPs).

High-Bandwidth Memory and Interconnect

Each MI455X unit is equipped with 432GB of HBM4 memory across 12 stacks, providing a peak memory bandwidth of 23.3 TB/s. This capacity and bandwidth are critical for supporting ever-larger AI models and context windows. Inter-accelerator communication is managed by 6th-generation Infinity Fabric, offering 3.6 TB/s of scale-up UALoE (Ultra Accelerator Link over Ethernet) bandwidth per GPU, and 600 GB/s peak scale-out bandwidth.

Instinct MI455X Key Specifications

Feature	Specification
Architecture	CDNA 5
Process Node	TSMC 2nm (XCDs) / 3nm (IOD, Fabric, Cache)
Transistors	320 Billion
Memory Capacity	432GB HBM4
Memory Bandwidth	23.3 TB/s Peak
FP8/FP16 Matrix Throughput	4.2x MI300X
Peak OCP MXFP4 Performance	40.3 PFLOPS
Peak OCP MXFP8 Performance	20.1 PFLOPS
Interconnect	6th-Gen Infinity Fabric, UALoE
Scale-Up Bandwidth (per GPU)	3.6 TB/s Bidirectional
TDP (Accelerator)	~850W – 1000W
Cooling	Liquid Cooling (Rack-scale)
Rack System	Helios Rack-Scale

Helios Rack-Scale Deployment and Thermal Management

The MI455X is integral to AMD’s new Helios rack-scale systems, designed for extreme density with 72 MI455X GPUs per rack. These clusters necessitate liquid cooling to manage the 850W-1000W TDP per accelerator. The Helios platform aims to directly compete with high-density systems like Nvidia’s NVL72 Blackwell B200 and Vera Rubin systems, offering competitive AI compute and superior HBM capacity.

ROCm 7.0 and Software Ecosystem

The launch coincides with ROCm 7.0, AMD’s open-source software platform. This release provides native PyTorch 2.5 zero-code migration, simplifying porting from CUDA environments. Support for OpenAI Triton kernels further enhances developer flexibility and performance optimization capabilities for custom AI workloads.

Key Technical Takeaways

• CDNA 5 architecture on TSMC 2nm/3nm delivers a 4.2x FP8/FP16 throughput increase over MI300X, featuring 320 billion transistors.
• HBM4 memory with 432GB and 23.3 TB/s bandwidth addresses growing model size and memory demands for frontier AI.
• 6th-gen Infinity Fabric with 3.6 TB/s scale-up UALoE bandwidth provides high-speed inter-GPU communication within Helios rack-scale systems.
• ROCm 7.0 with native PyTorch 2.5 and OpenAI Triton kernel support streamlines AI development and migration efforts.