Memory Model

Memory	Location, on/off chip	Cached	Access	Scope	Lifetime	Note
Register	on	n/a	R/W	1 thread	thread	no latency, no sharing, TB/s
Local	off	>=2.0	R/W	1 thread	thread
Shared	on	n/a	R/W	all threads in block	block	i.e: 64KB per block
Global	off	>=2.0	R/W	all threads + host	host allocation	GBs, cudamemcpy, cudamalloc
Constant	off	yes!	R	all threads + host	host allocation	1500GB/s
Texture	off	yes	R	all threads + host	host allocation

CUDA memory spaces & scopes

• global

• local (per-thread global memory)

• shared

• constant

• registers

CUDA memory spaces