Hello,
Just a small query: NVIDIAs GPUs are Multicore processors (many processors on a single chip), then why do we use the term Multiprocessor in connection with NVIDIA’s GPU arcitecture?
Thanks
just wanted to add more… what is the meaning of the word “streaming” in Streaming Multiprocessors (SM). Is this feature (“streaming”), nececsary for making the GPU, the GP-GPU (General Purpose GPU)?
you can read David Kanter’s article at
http://www.realworldtech.com/page.cfm?Arti…8195242&p=6
I think that you should define “core”, if “core” comes from CPU, for example, Intel Quadcore, then
each core is a complete execution unit, including instruction fetch, decode, logical/arithmetic operation, and write-back to
register file or external DRAM.
such definition of “core” corresponding to SM in GPU, not SP.
the best “mental model” presentation of GPU architecture I am aware of is by Kayvon Fatahalian, available here:
http://s09.idav.ucdavis.edu/talks/02_kayvo…uArchTalk09.pdf
It has some comparison of “marketing speak” vs. “what is actually in there” for GT200, RV700 and Larrabee which you might find useful. There’s also a corresponding paper: http://dx.doi.org/10.1145/1400181.1400197