Skip to main content

AMD MI455X

AMD Instinct MI455X is AMD's next-generation AI GPU announced at CES 2026, belonging to the Instinct MI400 series flagship model, built on CDNA 5 architecture with 2nm and 3nm hybrid process, equipped with 432GB HBM4 memory, delivering 40 PFLOPS FP4 compute, with performance 10× that of the previous-generation MI355X.

Key Specifications

SpecificationValue
GPU ArchitectureCDNA 5 architecture
Process Node2nm and 3nm hybrid (TSMC CoWoS-L packaging)
Transistor Count320 billion
GCD Configuration2 graphics compute dies (GCDs)
MCD Configuration2 memory controller dies (MCDs)
FP4 Compute40 PFLOPS
FP8 Compute20 PFLOPS
FP16 ComputeEstimated ~10 PFLOPS
FP32 ComputeEstimated ~5 TFLOPS
INT8 ComputeEstimated ~80 PFLOPS
Memory Capacity432 GB
Memory TypeHBM4
Memory Sites16 HBM4 memory sites
Memory Bandwidth19.6 TB/s
Interconnect5th-gen Infinity Fabric (3.6 TB/s)
TDPNot disclosed (estimated 1,500-2,000W)
Release DateJanuary 6, 2026
Mass ProductionSecond half of 2026
PricingNot disclosed

Architecture & Specifications

MI455X adopts a multi-chip module (MCM) design, with the core comprising:

  • 2 graphics compute die (GCD)
  • 2 memory controller die (MCD)
  • 16 HBM4 memory stacks

Key Technical Innovations

  1. CDNA 5 Architecture

    • Optimized for low-precision workloads
    • Supports FP4, FP8, and BF16 compute
    • Performance 10× MI355X
  2. 5th-Generation Infinity Fabric Interconnect

    • Single-node vertical scaling bandwidth up to 3.6 TB/s
    • Horizontal scaling bandwidth between nodes up to 300 GB/s
  3. HBM4 Memory

    • 432GB capacity (1.5× NVIDIA Rubin R200's 288GB)
    • 19.6 TB/s bandwidth (89% of NVIDIA Rubin R200's 22 TB/s)

Performance Comparison

ComparisonMI355XMI455XImprovement
FP4 Compute~4 PFLOPS40 PFLOPS10×
Memory Capacity288GB HBM3e432GB HBM41.5×
Memory Bandwidth~8 TB/s19.6 TB/s2.45×

MI400 Series Models

MI455X belongs to the AMD Instinct MI400 series, which also includes:

  1. MI455X: For large-scale AI training and inference (flagship)
  2. MI430X: Supports FP64 double-precision compute (HPC scenarios)
  3. MI440X: Designed for local enterprise AI (entry-level)

Companion Platform: Helios Rack

Helios Rack Configuration

  • 72× MI455X GPUs
  • 18× EPYC Venice CPUs (Zen 6, 256 cores)
  • Total Memory: 31 TB HBM4
  • Total Bandwidth: 1.4 PB/s
  • AI Inference Compute: 2.9 FP4 exaFLOPS
  • AI Training Compute: 1.4 FP8 exaFLOPS
  • TDP: ~300kW (full liquid cooling required)
  • Release Date: January 2026
  • Mass Production: Second half of 2026

Helios Supporting Hardware

  • CPU: AMD EPYC Venice (Zen 6, 2nm, 256 cores)
  • NIC: AMD Pensando Vulcano 800G AI NIC
  • Interconnect: Infinity Fabric, UALink, Ultra Ethernet
  • Network Adapters: Pensando Pollara 400G, Vulcano 800G

Mass Production & Delivery

  • Release Date: January 6, 2026 (CES 2026)
  • Mass Production: Second half of 2026
  • Helios System: Officially launched in second half of 2026
  • AMD Official Statement: Helios AI system development progress is on schedule, will be launched as planned in second half of 2026

Application Scenarios

MI455X targets large-scale AI training and distributed inference, suitable for:

  • Trillion-parameter LLM training
  • High-performance AI inference (low latency, high throughput)
  • Reinforcement learning (RL) and agentic AI
  • Scientific computing and HPC

References