Chiplet & MCM Architecture

MI300 vs Blackwell • Interactive Comparison • Hover for Details

AMD MI300X - 9-Die Chiplet

I/O DIE (IOD) 6nm • Infinity Fabric • PCIe Gen5 • HBM PHY Central hub for all 8 XCDs XCD0 38 CUs XCD1 38 CUs XCD2 38 CUs XCD3 38 CUs XCD4 38 CUs XCD5 38 CUs XCD6 38 CUs XCD7 38 CUs HBM0 24GB HBM1 24GB HBM2 24GB HBM3 24GB HBM4 24GB HBM5 24GB HBM6 24GB HBM7 24GB 2.5D CoWoS-S • Silicon Interposer • 9 Dies Total

NVIDIA Blackwell B200 - Dual-Die MCM

GPU DIE 0 104 SMs • 13,312 CUDA 4th Gen Tensor Cores 30MB L2 Cache TSMC 4nm • 104B transistors NVLink C2C 10 TB/s bidirectional GPU DIE 1 104 SMs • 13,312 CUDA 4th Gen Tensor Cores 30MB L2 Cache TSMC 4nm • 104B transistors NVLink 5.0 PCIe 6 / CXL CoWoS-L Package • 2 Dies • 208B Transistors Total

Specifications Comparison

AMD MI300X NVIDIA B200

Architecture

Die Configuration8 XCDs + 1 IOD2 GPU Dies
Process Node5nm (XCD) / 6nm (IOD)4nm
Transistors153 Billion208 Billion
PackagingCoWoS-S (2.5D)CoWoS-L (2.5D+)
Die-to-Die BWInfinity Fabric10 TB/s NVLink-C2C

Compute

Compute Units / SMs304 CUs208 SMs
Stream / CUDA Cores19,45626,624
Matrix Units304 MFMA832 Tensor Cores
FP16 Peak1,307 TFLOPS2,250 TFLOPS
FP8 Peak2,614 TFLOPS4,500 TFLOPS

Memory

HBM TypeHBM3HBM3e
Capacity192 GB192 GB
Bandwidth5.3 TB/s8.0 TB/s
L2 Cache256 MB60 MB

Interconnect & Power

Multi-GPUXGMI 896 GB/sNVLink 1.8 TB/s
Host InterfacePCIe Gen5PCIe Gen6 + CXL
TDP750W1000W
Max Scale8 GPUs (OAM)576 GPUs (NVL72)

Key Concepts

Chiplet
MCM (Multi-Chip Module)
2.5D Packaging
HBM (High Bandwidth Memory)
NVLink-C2C
Infinity Fabric
Silicon Interposer
TSV (Through-Silicon Via)
© 2026 Subramaniyam Pooni
CS²B Technologies