Blackwell SM Architecture

Streaming Multiprocessor (SM)

Tensor Core (tcgen05)

5th-gen MMA

CUDA Core

FP/INT units

Shared Memory (SMEM)

228 KB per SM

Tensor Memory (TMEM)

TMEM — 128 lanes

TMA Engine

Data Mover

SM ...

× N

Global Memory (GMEM)

Click a hardware unit to see details. Solid arrows = load path, dashed arrows = store path.