Why dont GPU have cache?

[ahhming]

if amd goal with HBM is to reduce latency & improve bandwidth, why not have a cache like a CPU do ?

Please enlighten me

[ZetZet]

Uhm gpu cache is the VRAM? You just need big amounts of cache for gpu's so they can't have super fast cache like cpu's do.

[Guest]

GPU's have registers

 

When the CPU writes to the register, the GPU goes off and does something.

[JefferyD90]

gpu's do have cache. Go look at pcper's coverage of the GTX 970 memory gate

[ChatDaw]

if amd goal with HBM is to reduce latency & improve bandwidth, why not have a cache like a CPU do ?

Please enlighten me

GPUs do have a small amount of cache.

[Casperian]

if amd goal with HBM is to reduce latency & improve bandwidth, why not have a cache like a CPU do ?

Please enlighten me

They do have cache, it's just spread out over a ton of cores. My 780 Ti (and all of Kepler) has 64KB of Cache per thread.

It is divided either 16/48, 32/32 or 48/16 into local and shared portions. The local is basically L1 cache and is used for the registers. L2 is for sharing between other threads in the block.

 

Look up the documentation on nVidia CUDA for a more in depth answer.

[MG2R]

Uhm gpu cache is the VRAM? You just need big amounts of cache for gpu's so they can't have super fast cache like cpu's do.

Cache is not the same as (V)RAM.

 

Cache is super-fast, super low-latency memory embedded onto the IC right next to the processing core(s) it's serving. VRAM is the same to a GPU as RAM is to a CPU. VRAM actually has a way higher bandwith than regular RAM. The tradeoff is that latency is bigger on VRAM.

[TheMcSame]

You might find this useful OP 

 

GTX 980 Die and block diagram.

 

[harrynowl]

check a block diagram and you will find GPUs have cache.

 

AMD GCN:

 

nvidia maxwell: