Skip to main content

AMD Instinct MI355X (288GB HBM3E)

Overview​

AMD Instinct MI355X is the 288GB HBM3E upgrade variant of the MI350 series, launching in H2 2025 (MI350 has 192GB HBM3E). Built on the CDNA 4 architecture, TSMC 3nm process, it features 288GB HBM3E memory (largest HBM capacity in the industry), 8 TB/s memory bandwidth (highest bandwidth in the industry), and 4.6 PFLOPS FP8 dense compute. It is AMD's key product between the NVIDIA B200 and MI400, targeting the AI large-model training and inference market.

Key upgrades (vs MI350X 192GB):

  • HBM capacity: 192GB → 288GB (+50%)
  • Memory bandwidth: 6.4 TB/s → 8 TB/s (+25%)
  • FP8 dense: 3.6 PF → 4.6 PF (+28%)
  • Supports FP4 / FP6 new precision formats (MI350 supports FP8 only)
  • Interconnect: UALoF (Ultra Accelerator Link Fabric) 600 GB/s (MI350 limited to 200 GB/s PCIe)

Core Specifications​

ItemSpecification
ArchitectureAMD CDNA 4 (same architecture as MI350)
Process NodeTSMC 3nm (N3)
GPU Cores304 CDNA 4 Compute Units
HBM288 GB HBM3E (largest in the industry)
HBM Channels8 stacks × 36GB HBM3E
Memory Bandwidth8 TB/s (highest in the industry)
FP4 sparse10.1 PFLOPS
FP6 sparse10.1 PFLOPS
FP8 dense5 PFLOPS
BF16 dense2.5 PFLOPS
FP16 dense2.5 PFLOPS
FP32115 TFLOPS
TDP~1400 W (typical board power)
Form FactorOAM / PCIe Gen5 ×16
InterconnectUALoF 600 GB/s (competes with NVLink 5)
Volume ProductionH2 2025
Unit Price (OAM)~$25,000 (estimated)

Comparison with MI350X 192GB​

MetricMI355X 288GBMI350X 192GBImprovement
Process Node3nm3nmSame
HBM Capacity288GB192GB+50%
HBM Bandwidth8 TB/s6.4 TB/s+25%
FP8 dense4.6 PF3.6 PF+28%
FP4 SupportYes (9.2 PF sparse)NoNew
FP6 SupportYes (6.9 PF sparse)NoNew
InterconnectUALoF 600 GB/sPCIe 5.0 200 GB/s3×
TDP750W750WSame
Price (estimated)~$25K~$20K+25%

Comparison with NVIDIA B200​

MetricAMD MI355XNVIDIA B200Difference
Memory288GB HBM3E192GB HBM3EMI355X +50%
Bandwidth8 TB/s8 TB/sSame
FP8 dense4.6 PF4.5 PF sparseMI355X slightly ahead (dense vs sparse)
FP4 sparse9.2 PF9 PF sparseSame
BF162.3 PF dense2.25 PF sparseMI355X slightly ahead
InterconnectUALoF 600 GB/sNVLink 5 1.8 TB/sB200 3×
TDP750W1000WMI355X -25%
SoftwareROCm 7 + OpenCUDA + ProprietaryAMD open
Price~$25K$30-40KMI355X -25%

MI355X advantages: Largest HBM capacity (288GB) + lowest TDP (750W) + open interconnect (UALoF) — one of the best hardware choices for large-model inference.

8 TB/s Memory Bandwidth Technology​

DimensionImplementation
HBM3E8 stacks × 1024-bit wide
Clock9.2 Gbps (industry's highest)
PHYAMD custom Infinity Fabric memory controller
PrefetchAdaptive prefetch algorithms
Error CorrectionOn-die ECC + Side-band ECC
DimensionSpecification
Bandwidth600 GB/s bidirectional
TopologyFully connected / Dragonfly+
ProtocolCustom (NVLink-like but open)
Latency~1 μs
SupportMI300X / MI325X / MI350X / MI355X / MI400 full product line
GovernanceUALink Consortium (established Q3 2024, AMD / Intel / Meta / Microsoft / Google and others)
2025 Members30+ companies
vs NVLink1/3 bandwidth, but fully open (NVLink is proprietary)

Strategic significance of UALoF: Break NVIDIA's NVLink monopoly. B200's 1.8 TB/s NVLink is 3× UALoF, but UALoF can interconnect any vendor's accelerators (NVIDIA / Groq / Habana / Tenstorrent), making it the future AI datacenter interconnect standard.

Vendor Information​

ItemDetails
CompanyAdvanced Micro Devices (AMD)
Product Pagehttps://www.amd.com/en/products/accelerators/instinct-mi350.html
CEOLisa Su
FoundryTSMC 3nm
H2 2025 Volume ProductionYes
2026 RoadmapMI400 (3nm+, 432GB HBM4)
2025 Revenue (MI business)~$8B (+80% YoY)
Key CustomersMicrosoft Azure (MAI platform), Meta, Oracle, Anthropic, Tenstorrent, newly launched LaminiAI

AMD Instinct Product Line​

ProductLaunchMemoryFP8 denseStatus
MI250XQ4 2021128GB HBM2E0 (FP16: 383 TF)EOL
MI300XQ4 2023192GB HBM31.3 PFIn production
MI325XQ4 2024256GB HBM3E2.6 PFIn production
MI350XQ3 2025192GB HBM3E3.6 PFIn production
MI355XH2 2025288GB HBM3E4.6 PFNew
MI4002026432GB HBM440 PF FP4 denseRoadmap

Key Features​

  • 288GB HBM3E: Industry's largest HBM capacity, exceeding NVIDIA B200's 192GB
  • 8 TB/s bandwidth: Industry's highest memory bandwidth
  • FP4 / FP6 / FP8 multi-precision: New low-precision support (same era as NVIDIA Blackwell)
  • UALoF 600 GB/s: Open interconnect, competing with NVLink
  • Helios rack: 72× MI355X + 36× EPYC Venice + Pensando NIC (H2 2025)
  • Open ROCm software: vs proprietary CUDA
  • Drawback: ROCm software maturity still 2-3 years behind CUDA

Helios Rack (72-GPU)​

ItemConfiguration
GPU Count72× MI355X
CPU Count36× EPYC Venice (256-core Zen 6)
NICPensando Vulcano 800GbE
GPU InterconnectUALoF fully connected
CPU-GPUPCIe Gen5 x16 + Infinity Fabric
Total Memory20.7 TB HBM3E
Total Compute331 PF FP8 dense
Rack TDP~80 kW
LaunchH2 2025 (alongside MI355X)

Use Cases​

  • ✅ Large model training (288GB accommodates larger models, UALoF multi-card interconnect)
  • ✅ LLM inference (288GB fits Llama 3 405B FP16 + large KV Cache)
  • ✅ Multimodal AI (Stable Diffusion 3, Sora training)
  • ✅ HPC + AI convergence (ROCm + MPI compatible)
  • ✅ Cloud providers (open ecosystem, multi-cloud deployment)
  • ✅ Government/state-owned enterprises (AMD US brand)
  • ❌ CUDA-only proprietary workloads
  • ❌ NVLink tightly-coupled code

MI355X vs MI400 (2026)​

MetricMI355X (H2 2025)MI400 (2026)Improvement
Memory288GB HBM3E432GB HBM4+50%
Bandwidth8 TB/s19.6 TB/s2.45×
FP4 dense4.6 PF FP840 PF FP4~9×
InterconnectUALoF 600 GB/sUALoF 1.3 TB/s2.2×
Process Node3nm3nm+ (N3P)Slightly newer
TDP750W~1000W+33%